Hugging Face's logo

Spaces:
cstr
/
LLMProviders
Running

App Files Community
CrispStrobe commited on
Commit
eed68bb
·
1 Parent(s): ac03708

feat: implement persistent HF private status and aggressive manual mappings for Qwen, DeepSeek, and Grok

Browse files
Files changed (3) hide show
  1. data/providers.json +825 -820
  2. scripts/fetch-providers.js +88 -22
  3. src/App.tsx +4 -0
data/providers.json CHANGED
@@ -998,22 +998,22 @@
998
  "eu_endpoints": true,
999
  "models": [
1000
  {
1001
- "name": "deepinfra/meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
1002
  "type": "chat",
1003
  "input_price_per_1m": 0.02,
1004
- "output_price_per_1m": 0.05,
1005
  "currency": "USD",
1006
- "context_window": 131072,
1007
- "size_b": 8
1008
  },
1009
  {
1010
- "name": "novita/meta-llama/llama-3.2-1b-instruct",
1011
  "type": "chat",
1012
  "input_price_per_1m": 0.02,
1013
- "output_price_per_1m": 0.02,
1014
  "currency": "USD",
1015
- "context_window": 131000,
1016
- "size_b": 1
1017
  },
1018
  {
1019
  "name": "openai/gpt-5-nano",
@@ -1051,44 +1051,32 @@
1051
  "hf_id": "meta-llama/Meta-Llama-3-8B-Instruct"
1052
  },
1053
  {
1054
- "name": "azure/gpt-5-nano",
1055
  "type": "vision",
1056
  "input_price_per_1m": 0.05,
1057
  "output_price_per_1m": 0.4,
1058
  "currency": "USD",
1059
  "capabilities": [
 
1060
  "reasoning"
1061
  ],
1062
- "context_window": 200000,
1063
- "hf_private": true
1064
- },
1065
- {
1066
- "name": "alibaba/qwen-turbo",
1067
- "type": "chat",
1068
- "input_price_per_1m": 0.05,
1069
- "output_price_per_1m": 0.2,
1070
- "currency": "USD",
1071
- "context_window": 1000000,
1072
- "capabilities": [
1073
- "tools"
1074
- ],
1075
  "hf_private": true
1076
  },
1077
  {
1078
- "name": "openai-responses/gpt-5-nano",
1079
  "type": "vision",
1080
  "input_price_per_1m": 0.05,
1081
  "output_price_per_1m": 0.4,
1082
  "currency": "USD",
1083
  "capabilities": [
1084
- "vision",
1085
  "reasoning"
1086
  ],
1087
- "context_window": 400000,
1088
  "hf_private": true
1089
  },
1090
  {
1091
- "name": "novita/Sao10K/L3-8B-Stheno-v3.2",
1092
  "type": "chat",
1093
  "input_price_per_1m": 0.05,
1094
  "output_price_per_1m": 0.05,
@@ -1110,7 +1098,7 @@
1110
  "hf_id": "meta-llama/Meta-Llama-3.1-8B-Instruct"
1111
  },
1112
  {
1113
- "name": "novita/sao10k/l3-8b-lunaris",
1114
  "type": "chat",
1115
  "input_price_per_1m": 0.05,
1116
  "output_price_per_1m": 0.05,
@@ -1118,6 +1106,18 @@
1118
  "context_window": 8192,
1119
  "size_b": 8
1120
  },
 
 
 
 
 
 
 
 
 
 
 
 
1121
  {
1122
  "name": "together/meta-llama/Llama-3.2-3B-Instruct-Turbo",
1123
  "type": "chat",
@@ -1128,22 +1128,22 @@
1128
  "size_b": 3
1129
  },
1130
  {
1131
- "name": "deepinfra/Qwen/Qwen2.5-Coder-32B-Instruct",
1132
  "type": "chat",
1133
  "input_price_per_1m": 0.07,
1134
- "output_price_per_1m": 0.16,
1135
  "currency": "USD",
1136
  "context_window": 16384,
1137
- "size_b": 32
1138
  },
1139
  {
1140
- "name": "deepinfra/microsoft/phi-4",
1141
  "type": "chat",
1142
  "input_price_per_1m": 0.07,
1143
- "output_price_per_1m": 0.14,
1144
  "currency": "USD",
1145
  "context_window": 16384,
1146
- "hf_private": true
1147
  },
1148
  {
1149
  "name": "novita/gryphe/mythomax-l2-13b",
@@ -1155,62 +1155,38 @@
1155
  "size_b": 13
1156
  },
1157
  {
1158
- "name": "azure/openai-responses/gpt-4.1-nano",
1159
- "type": "vision",
1160
- "input_price_per_1m": 0.1,
1161
- "output_price_per_1m": 0.4,
1162
- "currency": "USD",
1163
- "capabilities": [
1164
- "vision"
1165
- ],
1166
- "context_window": 1047576,
1167
- "hf_private": true
1168
- },
1169
- {
1170
- "name": "azure/gpt-4.1-nano",
1171
- "type": "vision",
1172
  "input_price_per_1m": 0.1,
1173
- "output_price_per_1m": 0.4,
1174
  "currency": "USD",
 
1175
  "capabilities": [
1176
- "vision"
1177
  ],
1178
- "context_window": 1047576,
1179
- "hf_private": true
1180
- },
1181
- {
1182
- "name": "together/meta-llama/Meta-Llama-3-8B-Instruct-Lite",
1183
- "type": "chat",
1184
- "input_price_per_1m": 0.1,
1185
- "output_price_per_1m": 0.1,
1186
- "currency": "USD",
1187
- "context_window": 8192,
1188
- "size_b": 8
1189
  },
1190
  {
1191
- "name": "deepinfra/Qwen/Qwen3-32B",
1192
  "type": "chat",
1193
  "input_price_per_1m": 0.1,
1194
  "output_price_per_1m": 0.3,
1195
  "currency": "USD",
1196
- "context_window": 40000,
1197
- "size_b": 32,
1198
  "capabilities": [
1199
- "tools",
1200
- "reasoning"
1201
  ],
1202
- "hf_id": "Qwen/Qwen3-32B"
 
1203
  },
1204
  {
1205
- "name": "openai-responses/gpt-4.1-nano",
1206
- "type": "vision",
1207
  "input_price_per_1m": 0.1,
1208
- "output_price_per_1m": 0.4,
1209
  "currency": "USD",
1210
- "capabilities": [
1211
- "vision"
1212
- ],
1213
- "context_window": 1047576,
1214
  "hf_private": true
1215
  },
1216
  {
@@ -1264,53 +1240,77 @@
1264
  "hf_private": true
1265
  },
1266
  {
1267
- "name": "mistral/devstral-small-latest",
1268
  "type": "chat",
1269
  "input_price_per_1m": 0.1,
1270
- "output_price_per_1m": 0.3,
 
 
 
 
 
 
 
 
 
1271
  "currency": "USD",
1272
  "context_window": 131072,
 
1273
  "capabilities": [
1274
- "tools"
 
1275
  ],
1276
- "hf_id": "mistralai/Devstral-Small-2507",
1277
- "size_b": 24
1278
  },
1279
  {
1280
- "name": "mistral/mistral-small-2503",
1281
- "type": "chat",
1282
  "input_price_per_1m": 0.1,
1283
- "output_price_per_1m": 0.3,
1284
  "currency": "USD",
1285
- "context_window": 32768,
 
 
 
1286
  "hf_private": true
1287
  },
1288
  {
1289
- "name": "mistral/devstral-small-2507",
1290
- "type": "chat",
1291
  "input_price_per_1m": 0.1,
1292
- "output_price_per_1m": 0.3,
1293
  "currency": "USD",
1294
- "context_window": 131072,
1295
  "capabilities": [
1296
- "tools"
1297
  ],
1298
- "hf_id": "mistralai/Devstral-Small-2507",
1299
- "size_b": 24
1300
  },
1301
  {
1302
- "name": "groq/openai/gpt-oss-20b",
 
 
 
 
 
 
 
 
 
 
 
 
1303
  "type": "chat",
1304
  "input_price_per_1m": 0.1,
1305
- "output_price_per_1m": 0.5,
1306
  "currency": "USD",
1307
- "context_window": 131072,
1308
- "size_b": 20,
1309
  "capabilities": [
1310
  "tools",
1311
  "reasoning"
1312
  ],
1313
- "hf_id": "openai/gpt-oss-20b"
1314
  },
1315
  {
1316
  "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct-Turbo",
@@ -1361,79 +1361,56 @@
1361
  "size_b": 8
1362
  },
1363
  {
1364
- "name": "parasail/parasail-qwen3-235b-a22b-instruct-2507",
1365
- "type": "chat",
1366
- "input_price_per_1m": 0.15,
1367
- "output_price_per_1m": 0.85,
1368
- "currency": "USD",
1369
- "context_window": 262144,
1370
- "size_b": 235
1371
- },
1372
- {
1373
- "name": "openai/gpt-4o-mini",
1374
  "type": "vision",
1375
  "input_price_per_1m": 0.15,
1376
  "output_price_per_1m": 0.6,
1377
  "currency": "USD",
1378
  "capabilities": [
1379
- "vision"
1380
- ],
1381
- "context_window": 128000,
1382
- "hf_private": true
1383
- },
1384
- {
1385
- "name": "novita/deepseek/deepseek-r1-distill-qwen-14b",
1386
- "type": "chat",
1387
- "input_price_per_1m": 0.15,
1388
- "output_price_per_1m": 0.15,
1389
- "currency": "USD",
1390
- "context_window": 128000,
1391
- "size_b": 14,
1392
- "capabilities": [
1393
- "tools",
1394
  "reasoning"
1395
  ],
1396
- "hf_id": "deepseek-ai/DeepSeek-R1"
 
1397
  },
1398
  {
1399
- "name": "nebius/openai/gpt-oss-120b",
1400
- "type": "chat",
1401
  "input_price_per_1m": 0.15,
1402
  "output_price_per_1m": 0.6,
1403
  "currency": "USD",
1404
  "capabilities": [
 
1405
  "reasoning"
1406
  ],
1407
- "context_window": 131000,
1408
- "size_b": 120,
1409
- "hf_id": "openai/gpt-oss-120b"
1410
  },
1411
  {
1412
- "name": "mistral/mistral-small-latest",
1413
  "type": "vision",
1414
  "input_price_per_1m": 0.15,
1415
  "output_price_per_1m": 0.6,
1416
  "currency": "USD",
1417
  "capabilities": [
1418
- "vision",
1419
- "reasoning"
1420
  ],
1421
- "context_window": 256000,
1422
  "hf_private": true
1423
  },
1424
  {
1425
- "name": "mistral/mistral-small-2603",
1426
- "type": "vision",
1427
  "input_price_per_1m": 0.15,
1428
  "output_price_per_1m": 0.6,
1429
  "currency": "USD",
1430
  "capabilities": [
1431
- "vision",
1432
  "reasoning"
1433
  ],
1434
- "context_window": 256000,
1435
- "size_b": 119,
1436
- "hf_id": "mistralai/Mistral-Small-4-119B-2603"
1437
  },
1438
  {
1439
  "name": "groq/openai/gpt-oss-120b",
@@ -1450,79 +1427,52 @@
1450
  "hf_id": "openai/gpt-oss-120b"
1451
  },
1452
  {
1453
- "name": "novita/mistralai/mistral-nemo",
1454
  "type": "chat",
1455
- "input_price_per_1m": 0.17,
1456
- "output_price_per_1m": 0.17,
1457
  "currency": "USD",
1458
- "context_window": 131072,
 
1459
  "capabilities": [
1460
- "tools"
 
1461
  ],
1462
- "hf_id": "mistralai/Mistral-Nemo-Instruct-2407",
1463
- "size_b": 12
1464
- },
1465
- {
1466
- "name": "together/meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
1467
- "type": "chat",
1468
- "input_price_per_1m": 0.18,
1469
- "output_price_per_1m": 0.18,
1470
- "currency": "USD",
1471
- "context_window": 131072,
1472
- "size_b": 8
1473
- },
1474
- {
1475
- "name": "together/meta-llama/LlamaGuard-2-8b",
1476
- "type": "chat",
1477
- "input_price_per_1m": 0.2,
1478
- "output_price_per_1m": 0.2,
1479
- "currency": "USD",
1480
- "context_window": 8192,
1481
- "size_b": 8
1482
  },
1483
  {
1484
- "name": "deepinfra/Qwen/Qwen3-235B-A22B",
1485
  "type": "chat",
1486
- "input_price_per_1m": 0.2,
1487
- "output_price_per_1m": 0.6,
1488
  "currency": "USD",
1489
- "context_window": 40000,
1490
- "size_b": 235,
1491
- "capabilities": [
1492
- "tools",
1493
- "reasoning"
1494
- ],
1495
- "hf_id": "Qwen/Qwen3-235B-A22B"
1496
  },
1497
  {
1498
- "name": "deepinfra/zai-org/GLM-4.5-Air",
1499
  "type": "chat",
1500
- "input_price_per_1m": 0.2,
1501
- "output_price_per_1m": 1.1,
1502
  "currency": "USD",
1503
  "context_window": 131072,
1504
  "capabilities": [
1505
- "tools",
1506
- "reasoning"
1507
  ],
1508
- "hf_id": "zai-org/GLM-4.5-Air",
1509
- "size_b": 110.5
1510
  },
1511
  {
1512
- "name": "alibaba/qwen3-30b-a3b-instruct-2507",
1513
- "type": "vision",
1514
- "input_price_per_1m": 0.2,
1515
- "output_price_per_1m": 0.8,
1516
  "currency": "USD",
1517
- "capabilities": [
1518
- "vision"
1519
- ],
1520
  "context_window": 131072,
1521
- "size_b": 30,
1522
- "hf_id": "Qwen/Qwen3-30B-A3B-Instruct-2507"
1523
  },
1524
  {
1525
- "name": "openai-responses/gpt-5.4-nano",
1526
  "type": "vision",
1527
  "input_price_per_1m": 0.2,
1528
  "output_price_per_1m": 1.25,
@@ -1535,7 +1485,16 @@
1535
  "hf_private": true
1536
  },
1537
  {
1538
- "name": "openai/gpt-5.4-nano",
 
 
 
 
 
 
 
 
 
1539
  "type": "vision",
1540
  "input_price_per_1m": 0.2,
1541
  "output_price_per_1m": 1.25,
@@ -1576,7 +1535,35 @@
1576
  "hf_id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
1577
  },
1578
  {
1579
- "name": "xai/grok-4-1-fast-non-reasoning",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1580
  "type": "vision",
1581
  "input_price_per_1m": 0.2,
1582
  "output_price_per_1m": 0.5,
@@ -1588,29 +1575,41 @@
1588
  "hf_private": true
1589
  },
1590
  {
1591
- "name": "xai/grok-4-fast",
1592
  "type": "vision",
1593
  "input_price_per_1m": 0.2,
1594
  "output_price_per_1m": 0.5,
1595
  "currency": "USD",
1596
- "context_window": 2000000,
1597
  "capabilities": [
1598
- "vision",
1599
- "tools",
1600
- "reasoning"
1601
  ],
 
1602
  "hf_private": true
1603
  },
1604
  {
1605
- "name": "xai/grok-4-1-fast-reasoning",
1606
  "type": "vision",
1607
  "input_price_per_1m": 0.2,
1608
- "output_price_per_1m": 0.5,
1609
  "currency": "USD",
1610
  "capabilities": [
1611
  "vision"
1612
  ],
 
 
 
 
 
 
 
 
 
1613
  "context_window": 2000000,
 
 
 
 
 
1614
  "hf_private": true
1615
  },
1616
  {
@@ -1628,29 +1627,30 @@
1628
  "hf_private": true
1629
  },
1630
  {
1631
- "name": "xai/grok-code-fast-1",
1632
  "type": "vision",
1633
  "input_price_per_1m": 0.2,
1634
- "output_price_per_1m": 1.5,
1635
  "currency": "USD",
1636
  "capabilities": [
1637
  "vision"
1638
  ],
1639
- "context_window": 256000,
1640
- "hf_private": true
 
1641
  },
1642
  {
1643
- "name": "deepinfra/Qwen/Qwen2.5-72B-Instruct",
1644
- "type": "vision",
1645
  "input_price_per_1m": 0.23,
1646
  "output_price_per_1m": 0.4,
1647
  "currency": "USD",
1648
  "context_window": 131072,
1649
- "size_b": 72,
1650
  "capabilities": [
1651
- "vision"
1652
  ],
1653
- "hf_id": "Qwen/Qwen2.5-VL-72B-Instruct"
1654
  },
1655
  {
1656
  "name": "deepinfra/deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
@@ -1666,54 +1666,50 @@
1666
  "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
1667
  },
1668
  {
1669
- "name": "deepinfra/meta-llama/Meta-Llama-3.1-70B-Instruct",
1670
- "type": "chat",
1671
  "input_price_per_1m": 0.23,
1672
  "output_price_per_1m": 0.4,
1673
  "currency": "USD",
1674
- "context_window": 130815,
1675
- "size_b": 70
 
 
 
 
1676
  },
1677
  {
1678
- "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct",
1679
  "type": "chat",
1680
  "input_price_per_1m": 0.23,
1681
  "output_price_per_1m": 0.4,
1682
  "currency": "USD",
1683
- "context_window": 131072,
1684
- "size_b": 70,
1685
- "capabilities": [
1686
- "tools"
1687
- ],
1688
- "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
1689
  },
1690
  {
1691
- "name": "azure/gpt-5-mini",
1692
- "type": "vision",
1693
  "input_price_per_1m": 0.25,
1694
- "output_price_per_1m": 2,
1695
  "currency": "USD",
1696
- "capabilities": [
1697
- "reasoning"
1698
- ],
1699
- "context_window": 200000,
1700
- "hf_private": true
1701
  },
1702
  {
1703
- "name": "openai-responses/gpt-5-mini",
1704
  "type": "vision",
1705
  "input_price_per_1m": 0.25,
1706
- "output_price_per_1m": 2,
1707
  "currency": "USD",
1708
  "capabilities": [
1709
- "vision",
1710
- "reasoning"
1711
  ],
1712
- "context_window": 400000,
1713
  "hf_private": true
1714
  },
1715
  {
1716
- "name": "vertex/gemini-3.1-flash-lite-preview",
1717
  "type": "vision",
1718
  "input_price_per_1m": 0.25,
1719
  "output_price_per_1m": 1.5,
@@ -1725,25 +1721,29 @@
1725
  "hf_private": true
1726
  },
1727
  {
1728
- "name": "google/gemini-3.1-flash-lite-preview",
1729
  "type": "vision",
1730
  "input_price_per_1m": 0.25,
1731
- "output_price_per_1m": 1.5,
1732
  "currency": "USD",
1733
  "capabilities": [
1734
- "vision"
 
1735
  ],
1736
- "context_window": 1048576,
1737
  "hf_private": true
1738
  },
1739
  {
1740
- "name": "mistral/open-mistral-7b",
1741
- "type": "chat",
1742
  "input_price_per_1m": 0.25,
1743
- "output_price_per_1m": 0.25,
1744
  "currency": "USD",
1745
- "context_window": 32768,
1746
- "size_b": 7
 
 
 
1747
  },
1748
  {
1749
  "name": "novita/deepseek/deepseek-v3.2",
@@ -1768,7 +1768,8 @@
1768
  "capabilities": [
1769
  "tools"
1770
  ],
1771
- "hf_id": "deepseek-ai/DeepSeek-V3"
 
1772
  },
1773
  {
1774
  "name": "deepseek/deepseek-reasoner",
@@ -1780,47 +1781,12 @@
1780
  "hf_private": true
1781
  },
1782
  {
1783
- "name": "together/Qwen/Qwen2.5-7B-Instruct-Turbo",
1784
- "type": "chat",
1785
- "input_price_per_1m": 0.3,
1786
- "output_price_per_1m": 0.3,
1787
- "currency": "USD",
1788
- "context_window": 32768,
1789
- "size_b": 7
1790
- },
1791
- {
1792
- "name": "deepinfra/deepseek-ai/DeepSeek-V3.1",
1793
- "type": "chat",
1794
- "input_price_per_1m": 0.3,
1795
- "output_price_per_1m": 1,
1796
- "currency": "USD",
1797
- "context_window": 163840,
1798
- "capabilities": [
1799
- "tools",
1800
- "reasoning"
1801
- ],
1802
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
1803
- "size_b": 671
1804
- },
1805
- {
1806
- "name": "parasail/parasail-gemma3-27b-it",
1807
  "type": "chat",
1808
  "input_price_per_1m": 0.3,
1809
- "output_price_per_1m": 0.5,
1810
- "currency": "USD",
1811
- "context_window": 128000,
1812
- "size_b": 27
1813
- },
1814
- {
1815
- "name": "alibaba/qwen3-coder-flash",
1816
- "type": "vision",
1817
- "input_price_per_1m": 0.3,
1818
- "output_price_per_1m": 1.5,
1819
  "currency": "USD",
1820
- "capabilities": [
1821
- "vision"
1822
- ],
1823
- "context_window": 1048576,
1824
  "hf_private": true
1825
  },
1826
  {
@@ -1850,7 +1816,7 @@
1850
  "hf_private": true
1851
  },
1852
  {
1853
- "name": "novita/minimax-m2.7",
1854
  "type": "vision",
1855
  "input_price_per_1m": 0.3,
1856
  "output_price_per_1m": 1.2,
@@ -1864,20 +1830,21 @@
1864
  "hf_private": true
1865
  },
1866
  {
1867
- "name": "novita/deepseek/deepseek-r1-distill-qwen-32b",
1868
  "type": "chat",
1869
  "input_price_per_1m": 0.3,
1870
- "output_price_per_1m": 0.3,
1871
  "currency": "USD",
1872
- "context_window": 12800,
1873
- "size_b": 32,
1874
  "capabilities": [
 
1875
  "reasoning"
1876
  ],
1877
- "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
 
1878
  },
1879
  {
1880
- "name": "novita/MiniMax-M2.7",
1881
  "type": "vision",
1882
  "input_price_per_1m": 0.3,
1883
  "output_price_per_1m": 1.2,
@@ -1887,23 +1854,11 @@
1887
  "reasoning"
1888
  ],
1889
  "context_window": 200000,
1890
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
1891
- "hf_private": true
1892
- },
1893
- {
1894
- "name": "xai/grok-3-mini",
1895
- "type": "chat",
1896
- "input_price_per_1m": 0.3,
1897
- "output_price_per_1m": 0.5,
1898
- "currency": "USD",
1899
- "capabilities": [
1900
- "reasoning"
1901
- ],
1902
- "context_window": 131072,
1903
- "hf_private": true
1904
  },
1905
  {
1906
- "name": "coding/gemini-2.5-flash",
1907
  "type": "vision",
1908
  "input_price_per_1m": 0.3,
1909
  "output_price_per_1m": 2.5,
@@ -1942,7 +1897,16 @@
1942
  "size_b": 685.4
1943
  },
1944
  {
1945
- "name": "minimaxi/MiniMax-M2.7",
 
 
 
 
 
 
 
 
 
1946
  "type": "vision",
1947
  "input_price_per_1m": 0.3,
1948
  "output_price_per_1m": 1.2,
@@ -1956,55 +1920,92 @@
1956
  "hf_private": true
1957
  },
1958
  {
1959
- "name": "minimaxi/MiniMax-M2",
1960
- "type": "chat",
1961
  "input_price_per_1m": 0.3,
1962
  "output_price_per_1m": 1.2,
1963
  "currency": "USD",
 
 
 
 
1964
  "context_window": 200000,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1965
  "capabilities": [
1966
  "tools",
1967
  "reasoning"
1968
  ],
1969
- "hf_id": "MiniMaxAI/MiniMax-M2",
1970
- "size_b": 228.7
 
 
 
 
 
 
 
 
 
1971
  },
1972
  {
1973
- "name": "minimaxi/MiniMax-M2.5",
1974
- "type": "vision",
1975
  "input_price_per_1m": 0.3,
1976
- "output_price_per_1m": 1.2,
1977
  "currency": "USD",
1978
  "capabilities": [
1979
- "vision",
1980
  "reasoning"
1981
  ],
1982
- "context_window": 200000,
1983
- "hf_id": "MiniMaxAI/MiniMax-M2.5",
1984
- "size_b": 228.7
1985
  },
1986
  {
1987
- "name": "google/gemini-2.5-flash",
1988
  "type": "vision",
1989
  "input_price_per_1m": 0.3,
1990
- "output_price_per_1m": 2.5,
1991
  "currency": "USD",
1992
  "capabilities": [
1993
- "vision",
1994
- "reasoning"
1995
  ],
1996
  "context_window": 1048576,
1997
  "hf_private": true
1998
  },
1999
- {
2000
- "name": "mistral/codestral-latest",
2001
- "type": "chat",
2002
- "input_price_per_1m": 0.3,
2003
- "output_price_per_1m": 0.9,
2004
- "currency": "USD",
2005
- "context_window": 131072,
2006
- "hf_private": true
2007
- },
2008
  {
2009
  "name": "deepinfra/meta-llama/Llama-3.2-90B-Vision-Instruct",
2010
  "type": "chat",
@@ -2041,19 +2042,37 @@
2041
  "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2042
  },
2043
  {
2044
- "name": "azure/gpt-4.1-mini",
2045
- "type": "vision",
2046
  "input_price_per_1m": 0.4,
2047
- "output_price_per_1m": 1.6,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2048
  "currency": "USD",
 
2049
  "capabilities": [
2050
- "vision"
2051
  ],
2052
- "context_window": 1047576,
2053
  "hf_private": true
2054
  },
2055
  {
2056
- "name": "azure/openai-responses/gpt-4.1-mini",
2057
  "type": "vision",
2058
  "input_price_per_1m": 0.4,
2059
  "output_price_per_1m": 1.6,
@@ -2065,12 +2084,12 @@
2065
  "hf_private": true
2066
  },
2067
  {
2068
- "name": "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct",
2069
  "type": "chat",
2070
  "input_price_per_1m": 0.4,
2071
- "output_price_per_1m": 1.6,
2072
  "currency": "USD",
2073
- "context_window": 262144,
2074
  "size_b": 480,
2075
  "capabilities": [
2076
  "tools"
@@ -2078,16 +2097,17 @@
2078
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2079
  },
2080
  {
2081
- "name": "alibaba/qwen-plus",
2082
  "type": "chat",
2083
  "input_price_per_1m": 0.4,
2084
- "output_price_per_1m": 1.2,
2085
  "currency": "USD",
2086
- "context_window": 131072,
2087
  "capabilities": [
2088
- "tools"
2089
  ],
2090
- "hf_private": true
 
 
2091
  },
2092
  {
2093
  "name": "openai-responses/gpt-4.1-mini",
@@ -2102,7 +2122,19 @@
2102
  "hf_private": true
2103
  },
2104
  {
2105
- "name": "openai/gpt-4.1-mini",
 
 
 
 
 
 
 
 
 
 
 
 
2106
  "type": "vision",
2107
  "input_price_per_1m": 0.4,
2108
  "output_price_per_1m": 1.6,
@@ -2137,25 +2169,12 @@
2137
  "hf_private": true
2138
  },
2139
  {
2140
- "name": "nebius/zai-org/GLM-4.7",
2141
- "type": "chat",
2142
- "input_price_per_1m": 0.4,
2143
- "output_price_per_1m": 2,
2144
- "currency": "USD",
2145
- "capabilities": [
2146
- "reasoning"
2147
- ],
2148
- "context_window": 200000,
2149
- "hf_id": "zai-org/GLM-4.7",
2150
- "size_b": 358.3
2151
- },
2152
- {
2153
- "name": "nebius/Qwen/Qwen3-Coder-480B-A35B-Instruct",
2154
  "type": "chat",
2155
  "input_price_per_1m": 0.4,
2156
- "output_price_per_1m": 1.8,
2157
  "currency": "USD",
2158
- "context_window": 262000,
2159
  "size_b": 480,
2160
  "capabilities": [
2161
  "tools"
@@ -2163,19 +2182,22 @@
2163
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2164
  },
2165
  {
2166
- "name": "mistral/mistral-medium-latest",
2167
  "type": "chat",
2168
  "input_price_per_1m": 0.4,
2169
- "output_price_per_1m": 2,
2170
  "currency": "USD",
2171
  "context_window": 131072,
 
 
 
2172
  "hf_private": true
2173
  },
2174
  {
2175
- "name": "mistral/devstral-medium-2507",
2176
  "type": "chat",
2177
- "input_price_per_1m": 0.4,
2178
- "output_price_per_1m": 2,
2179
  "currency": "USD",
2180
  "context_window": 131072,
2181
  "capabilities": [
@@ -2183,15 +2205,6 @@
2183
  ],
2184
  "hf_private": true
2185
  },
2186
- {
2187
- "name": "mistral/devstral-latest",
2188
- "type": "chat",
2189
- "input_price_per_1m": 0.4,
2190
- "output_price_per_1m": 2,
2191
- "currency": "USD",
2192
- "context_window": 256000,
2193
- "hf_private": true
2194
- },
2195
  {
2196
  "name": "vertex/gemini-3.1-flash-image-preview",
2197
  "type": "vision",
@@ -2218,47 +2231,6 @@
2218
  "context_window": 1048576,
2219
  "hf_private": true
2220
  },
2221
- {
2222
- "name": "nebius/moonshotai/Kimi-K2-Instruct",
2223
- "type": "chat",
2224
- "input_price_per_1m": 0.5,
2225
- "output_price_per_1m": 2.4,
2226
- "currency": "USD",
2227
- "context_window": 131000,
2228
- "capabilities": [
2229
- "tools"
2230
- ],
2231
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2232
- "size_b": 1026.5
2233
- },
2234
- {
2235
- "name": "nebius/moonshotai/kimi-k2.5",
2236
- "type": "vision",
2237
- "input_price_per_1m": 0.5,
2238
- "output_price_per_1m": 2.5,
2239
- "currency": "USD",
2240
- "capabilities": [
2241
- "vision",
2242
- "reasoning"
2243
- ],
2244
- "context_window": 256000,
2245
- "hf_id": "moonshotai/Kimi-K2.5",
2246
- "size_b": 1058.6
2247
- },
2248
- {
2249
- "name": "nebius/deepseek-ai/DeepSeek-V3-0324",
2250
- "type": "chat",
2251
- "input_price_per_1m": 0.5,
2252
- "output_price_per_1m": 1.5,
2253
- "currency": "USD",
2254
- "context_window": 128000,
2255
- "capabilities": [
2256
- "tools",
2257
- "reasoning"
2258
- ],
2259
- "size_b": 685,
2260
- "hf_id": "deepseek-ai/DeepSeek-V3-0324"
2261
- },
2262
  {
2263
  "name": "google/gemini-3.1-flash-image-preview",
2264
  "type": "vision",
@@ -2286,138 +2258,111 @@
2286
  "hf_private": true
2287
  },
2288
  {
2289
- "name": "mistral/mistral-large-latest",
2290
- "type": "chat",
2291
  "input_price_per_1m": 0.5,
2292
- "output_price_per_1m": 1.5,
2293
  "currency": "USD",
2294
- "context_window": 131072,
2295
  "capabilities": [
2296
- "tools"
 
2297
  ],
2298
- "hf_private": true
 
 
2299
  },
2300
  {
2301
- "name": "novita/meta-llama/llama-3-70b-instruct",
2302
  "type": "chat",
2303
- "input_price_per_1m": 0.51,
2304
- "output_price_per_1m": 0.74,
2305
  "currency": "USD",
2306
- "context_window": 8192,
2307
- "size_b": 70,
2308
  "capabilities": [
2309
  "tools"
2310
  ],
2311
- "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2312
- },
2313
- {
2314
- "name": "openai/o4-mini",
2315
- "type": "vision",
2316
- "input_price_per_1m": 0.55,
2317
- "output_price_per_1m": 2.2,
2318
- "currency": "USD",
2319
- "capabilities": [
2320
- "reasoning"
2321
- ],
2322
- "context_window": 200000,
2323
- "hf_private": true
2324
  },
2325
  {
2326
- "name": "vertex/deepseek-v3.2",
2327
- "type": "vision",
2328
- "input_price_per_1m": 0.56,
2329
- "output_price_per_1m": 1.68,
2330
  "currency": "USD",
 
2331
  "capabilities": [
2332
- "vision",
2333
  "reasoning"
2334
  ],
2335
- "context_window": 163840,
2336
- "hf_id": "deepseek-ai/DeepSeek-V3.2",
2337
- "size_b": 685.4
2338
  },
2339
  {
2340
- "name": "novita/moonshotai/kimi-k2-instruct",
2341
  "type": "chat",
2342
- "input_price_per_1m": 0.57,
2343
- "output_price_per_1m": 2.3,
2344
  "currency": "USD",
2345
- "context_window": 131072,
 
2346
  "capabilities": [
2347
  "tools"
2348
  ],
2349
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2350
- "size_b": 1026.5
2351
- },
2352
- {
2353
- "name": "deepinfra/zai-org/GLM-4.5",
2354
- "type": "chat",
2355
- "input_price_per_1m": 0.6,
2356
- "output_price_per_1m": 2.2,
2357
- "currency": "USD",
2358
- "context_window": 131072,
2359
- "capabilities": [
2360
- "tools",
2361
- "reasoning"
2362
- ],
2363
- "hf_id": "zai-org/GLM-4.5",
2364
- "size_b": 358.3
2365
  },
2366
  {
2367
- "name": "alibaba/qwen3.5",
2368
  "type": "vision",
2369
- "input_price_per_1m": 0.6,
2370
- "output_price_per_1m": 3.6,
2371
  "currency": "USD",
2372
  "capabilities": [
2373
- "vision",
2374
  "reasoning"
2375
  ],
2376
- "context_window": 256000,
2377
- "size_b": 9,
2378
- "hf_id": "Qwen/Qwen3.5-9B"
2379
  },
2380
  {
2381
- "name": "novita/zai-org/glm-4.5",
2382
- "type": "chat",
2383
- "input_price_per_1m": 0.6,
2384
- "output_price_per_1m": 2.2,
2385
  "currency": "USD",
2386
- "context_window": 131072,
2387
  "capabilities": [
2388
- "tools",
2389
  "reasoning"
2390
  ],
2391
- "hf_id": "zai-org/GLM-4.5",
2392
- "size_b": 358.3
 
2393
  },
2394
  {
2395
- "name": "novita/zai-org/glm-4.6",
2396
  "type": "chat",
2397
- "input_price_per_1m": 0.6,
2398
- "output_price_per_1m": 2.2,
2399
  "currency": "USD",
2400
- "context_window": 204800,
2401
  "capabilities": [
2402
- "tools",
2403
- "reasoning"
2404
  ],
2405
- "hf_id": "zai-org/GLM-4.6",
2406
- "size_b": 356.8
2407
  },
2408
  {
2409
- "name": "nebius/zai-org/GLM-4.5",
2410
- "type": "chat",
2411
  "input_price_per_1m": 0.6,
2412
- "output_price_per_1m": 2.2,
2413
  "currency": "USD",
2414
- "context_window": 128000,
2415
  "capabilities": [
2416
- "tools",
2417
  "reasoning"
2418
  ],
2419
- "hf_id": "zai-org/GLM-4.5",
2420
- "size_b": 358.3
 
2421
  },
2422
  {
2423
  "name": "minimaxi/MiniMax-M2.5-highspeed",
@@ -2434,18 +2379,31 @@
2434
  "size_b": 228.7
2435
  },
2436
  {
2437
- "name": "minimaxi/MiniMax-M2.7-highspeed",
 
 
 
 
 
 
 
 
 
 
 
 
 
2438
  "type": "vision",
2439
  "input_price_per_1m": 0.6,
2440
- "output_price_per_1m": 2.4,
2441
  "currency": "USD",
2442
  "capabilities": [
2443
  "vision",
2444
  "reasoning"
2445
  ],
2446
- "context_window": 200000,
2447
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
2448
- "hf_private": true
2449
  },
2450
  {
2451
  "name": "moonshot/kimi-k2-thinking-turbo",
@@ -2488,31 +2446,18 @@
2488
  "size_b": 1026.5
2489
  },
2490
  {
2491
- "name": "moonshot/kimi-k2-0711-preview",
2492
  "type": "chat",
2493
  "input_price_per_1m": 0.6,
2494
- "output_price_per_1m": 2.5,
2495
- "currency": "USD",
2496
- "context_window": 131072,
2497
- "capabilities": [
2498
- "tools"
2499
- ],
2500
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2501
- "size_b": 1026.5
2502
- },
2503
- {
2504
- "name": "moonshot/kimi-k2.5",
2505
- "type": "vision",
2506
- "input_price_per_1m": 0.6,
2507
- "output_price_per_1m": 3,
2508
  "currency": "USD",
 
2509
  "capabilities": [
2510
- "vision",
2511
  "reasoning"
2512
  ],
2513
- "context_window": 262144,
2514
- "hf_id": "moonshotai/Kimi-K2.5",
2515
- "size_b": 1058.6
2516
  },
2517
  {
2518
  "name": "zai/GLM-4.6",
@@ -2553,6 +2498,62 @@
2553
  "hf_id": "zai-org/GLM-4.5",
2554
  "size_b": 358.3
2555
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2556
  {
2557
  "name": "novita/microsoft/wizardlm-2-8x22b",
2558
  "type": "chat",
@@ -2576,16 +2577,18 @@
2576
  "hf_private": true
2577
  },
2578
  {
2579
- "name": "parasail/parasail-qwen25-vl-72b-instruct",
2580
- "type": "vision",
2581
  "input_price_per_1m": 0.7,
2582
- "output_price_per_1m": 0.7,
2583
  "currency": "USD",
 
2584
  "capabilities": [
2585
- "vision"
 
2586
  ],
2587
- "context_window": 32768,
2588
- "size_b": 72
2589
  },
2590
  {
2591
  "name": "novita/deepseek/deepseek-prover-v2-671b",
@@ -2597,21 +2600,19 @@
2597
  "size_b": 671
2598
  },
2599
  {
2600
- "name": "novita/deepseek/deepseek-r1-turbo",
2601
- "type": "chat",
2602
  "input_price_per_1m": 0.7,
2603
- "output_price_per_1m": 2.5,
2604
  "currency": "USD",
2605
- "context_window": 64000,
2606
  "capabilities": [
2607
- "tools",
2608
- "reasoning"
2609
  ],
2610
- "hf_id": "deepseek-ai/DeepSeek-R1",
2611
- "size_b": 684.5
2612
  },
2613
  {
2614
- "name": "openai-responses/gpt-5.4-mini",
2615
  "type": "vision",
2616
  "input_price_per_1m": 0.75,
2617
  "output_price_per_1m": 4.5,
@@ -2624,7 +2625,7 @@
2624
  "hf_private": true
2625
  },
2626
  {
2627
- "name": "openai/gpt-5.4-mini",
2628
  "type": "vision",
2629
  "input_price_per_1m": 0.75,
2630
  "output_price_per_1m": 4.5,
@@ -2637,13 +2638,18 @@
2637
  "hf_private": true
2638
  },
2639
  {
2640
- "name": "deepinfra/meta-llama/Meta-Llama-3.1-405B-Instruct",
2641
  "type": "chat",
2642
  "input_price_per_1m": 0.8,
2643
- "output_price_per_1m": 0.8,
2644
  "currency": "USD",
2645
- "context_window": 130815,
2646
- "size_b": 405
 
 
 
 
 
2647
  },
2648
  {
2649
  "name": "novita/deepseek/deepseek-r1-distill-llama-70b",
@@ -2672,18 +2678,13 @@
2672
  "hf_id": "Qwen/Qwen2.5-VL-72B-Instruct"
2673
  },
2674
  {
2675
- "name": "nebius/deepseek-ai/DeepSeek-R1-0528",
2676
  "type": "chat",
2677
  "input_price_per_1m": 0.8,
2678
- "output_price_per_1m": 2.4,
2679
  "currency": "USD",
2680
- "context_window": 164000,
2681
- "capabilities": [
2682
- "tools",
2683
- "reasoning"
2684
- ],
2685
- "hf_id": "deepseek-ai/DeepSeek-R1-0528",
2686
- "size_b": 684.5
2687
  },
2688
  {
2689
  "name": "deepinfra/deepseek-ai/DeepSeek-R1",
@@ -2697,7 +2698,7 @@
2697
  "reasoning"
2698
  ],
2699
  "hf_id": "deepseek-ai/DeepSeek-R1",
2700
- "size_b": 684.5
2701
  },
2702
  {
2703
  "name": "deepinfra/deepseek-ai/DeepSeek-V3",
@@ -2759,51 +2760,15 @@
2759
  "reasoning"
2760
  ],
2761
  "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
2762
- "size_b": 671
2763
- },
2764
- {
2765
- "name": "parasail/parasail-kimi-k2-instruct",
2766
- "type": "chat",
2767
- "input_price_per_1m": 0.99,
2768
- "output_price_per_1m": 2.99,
2769
- "currency": "USD",
2770
- "context_window": 131072,
2771
- "hf_private": true
2772
- },
2773
- {
2774
- "name": "anthropic/claude-haiku-4-5",
2775
- "type": "vision",
2776
- "input_price_per_1m": 1,
2777
- "output_price_per_1m": 5,
2778
- "currency": "USD",
2779
- "capabilities": [
2780
- "vision"
2781
- ],
2782
- "context_window": 200000,
2783
- "hf_private": true
2784
- },
2785
- {
2786
- "name": "alibaba/qwen3-coder-plus",
2787
- "type": "vision",
2788
- "input_price_per_1m": 1,
2789
- "output_price_per_1m": 5,
2790
- "currency": "USD",
2791
- "capabilities": [
2792
- "vision"
2793
- ],
2794
- "context_window": 1048576,
2795
- "hf_private": true
2796
  },
2797
  {
2798
- "name": "perplexity/sonar",
2799
- "type": "vision",
2800
- "input_price_per_1m": 1,
2801
- "output_price_per_1m": 1,
2802
  "currency": "USD",
2803
  "context_window": 131072,
2804
- "capabilities": [
2805
- "vision"
2806
- ],
2807
  "hf_private": true
2808
  },
2809
  {
@@ -2832,20 +2797,7 @@
2832
  "hf_private": true
2833
  },
2834
  {
2835
- "name": "novita/GLM-5",
2836
- "type": "chat",
2837
- "input_price_per_1m": 1,
2838
- "output_price_per_1m": 3.2,
2839
- "currency": "USD",
2840
- "capabilities": [
2841
- "reasoning"
2842
- ],
2843
- "context_window": 202800,
2844
- "hf_id": "zai-org/GLM-5",
2845
- "size_b": 753.9
2846
- },
2847
- {
2848
- "name": "bedrock/claude-haiku-4-5",
2849
  "type": "vision",
2850
  "input_price_per_1m": 1,
2851
  "output_price_per_1m": 5,
@@ -2897,19 +2849,56 @@
2897
  "size_b": 1026.5
2898
  },
2899
  {
2900
- "name": "azure/o4-mini",
2901
- "type": "vision",
2902
- "input_price_per_1m": 1.1,
2903
- "output_price_per_1m": 4.4,
2904
  "currency": "USD",
2905
  "capabilities": [
2906
  "reasoning"
2907
  ],
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2908
  "context_window": 200000,
2909
  "hf_private": true
2910
  },
2911
  {
2912
- "name": "openai-responses/o3-mini",
 
 
 
 
 
 
 
 
 
 
 
 
2913
  "type": "chat",
2914
  "input_price_per_1m": 1.1,
2915
  "output_price_per_1m": 4.4,
@@ -2933,7 +2922,7 @@
2933
  "hf_private": true
2934
  },
2935
  {
2936
- "name": "openai/o3-mini",
2937
  "type": "chat",
2938
  "input_price_per_1m": 1.1,
2939
  "output_price_per_1m": 4.4,
@@ -2945,13 +2934,16 @@
2945
  "hf_private": true
2946
  },
2947
  {
2948
- "name": "together/Qwen/Qwen2.5-72B-Instruct-Turbo",
2949
- "type": "chat",
2950
- "input_price_per_1m": 1.2,
2951
- "output_price_per_1m": 1.2,
2952
  "currency": "USD",
2953
- "context_window": 32768,
2954
- "size_b": 72
 
 
 
2955
  },
2956
  {
2957
  "name": "moonshot/kimi-k2-turbo-preview",
@@ -2967,45 +2959,55 @@
2967
  "size_b": 1026.5
2968
  },
2969
  {
2970
- "name": "azure/gpt-5",
 
 
 
 
 
 
 
 
 
2971
  "type": "vision",
2972
  "input_price_per_1m": 1.25,
2973
  "output_price_per_1m": 10,
2974
  "currency": "USD",
2975
  "capabilities": [
 
2976
  "reasoning"
2977
  ],
2978
- "context_window": 200000,
2979
  "hf_private": true
2980
  },
2981
  {
2982
- "name": "azure/gpt-5.1",
2983
  "type": "vision",
2984
  "input_price_per_1m": 1.25,
2985
  "output_price_per_1m": 10,
2986
  "currency": "USD",
2987
  "capabilities": [
 
2988
  "reasoning"
2989
  ],
2990
- "context_window": 200000,
2991
  "hf_private": true
2992
  },
2993
  {
2994
- "name": "together/deepseek-ai/DeepSeek-V3",
2995
- "type": "chat",
2996
  "input_price_per_1m": 1.25,
2997
- "output_price_per_1m": 1.25,
2998
  "currency": "USD",
2999
- "context_window": 131072,
3000
  "capabilities": [
3001
- "tools",
3002
  "reasoning"
3003
  ],
3004
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
3005
- "size_b": 671
3006
  },
3007
  {
3008
- "name": "openai-responses/gpt-5.1",
3009
  "type": "vision",
3010
  "input_price_per_1m": 1.25,
3011
  "output_price_per_1m": 10,
@@ -3018,7 +3020,7 @@
3018
  "hf_private": true
3019
  },
3020
  {
3021
- "name": "openai-responses/gpt-5-codex",
3022
  "type": "vision",
3023
  "input_price_per_1m": 1.25,
3024
  "output_price_per_1m": 10,
@@ -3027,24 +3029,25 @@
3027
  "vision",
3028
  "reasoning"
3029
  ],
3030
- "context_window": 400000,
3031
  "hf_private": true
3032
  },
3033
  {
3034
- "name": "openai-responses/gpt-5.1-codex",
3035
- "type": "vision",
3036
  "input_price_per_1m": 1.25,
3037
- "output_price_per_1m": 10,
3038
  "currency": "USD",
 
3039
  "capabilities": [
3040
- "vision",
3041
  "reasoning"
3042
  ],
3043
- "context_window": 400000,
3044
- "hf_private": true
3045
  },
3046
  {
3047
- "name": "openai-responses/gpt-5",
3048
  "type": "vision",
3049
  "input_price_per_1m": 1.25,
3050
  "output_price_per_1m": 10,
@@ -3057,7 +3060,7 @@
3057
  "hf_private": true
3058
  },
3059
  {
3060
- "name": "vertex/gemini-2.5-pro",
3061
  "type": "vision",
3062
  "input_price_per_1m": 1.25,
3063
  "output_price_per_1m": 10,
@@ -3066,11 +3069,11 @@
3066
  "vision",
3067
  "reasoning"
3068
  ],
3069
- "context_window": 1048576,
3070
  "hf_private": true
3071
  },
3072
  {
3073
- "name": "openai/gpt-5.1",
3074
  "type": "vision",
3075
  "input_price_per_1m": 1.25,
3076
  "output_price_per_1m": 10,
@@ -3083,7 +3086,7 @@
3083
  "hf_private": true
3084
  },
3085
  {
3086
- "name": "openai/gpt-5.1-chat",
3087
  "type": "vision",
3088
  "input_price_per_1m": 1.25,
3089
  "output_price_per_1m": 10,
@@ -3092,37 +3095,35 @@
3092
  "vision",
3093
  "reasoning"
3094
  ],
3095
- "context_window": 128000,
3096
  "hf_private": true
3097
  },
3098
  {
3099
- "name": "openai/gpt-5-chat",
3100
  "type": "vision",
3101
  "input_price_per_1m": 1.25,
3102
  "output_price_per_1m": 10,
3103
  "currency": "USD",
3104
  "capabilities": [
3105
- "vision",
3106
  "reasoning"
3107
  ],
3108
- "context_window": 128000,
3109
  "hf_private": true
3110
  },
3111
  {
3112
- "name": "coding/gemini-2.5-pro",
3113
  "type": "vision",
3114
  "input_price_per_1m": 1.25,
3115
  "output_price_per_1m": 10,
3116
  "currency": "USD",
3117
  "capabilities": [
3118
- "vision",
3119
  "reasoning"
3120
  ],
3121
- "context_window": 1048576,
3122
  "hf_private": true
3123
  },
3124
  {
3125
- "name": "google/gemini-2.5-pro",
3126
  "type": "vision",
3127
  "input_price_per_1m": 1.25,
3128
  "output_price_per_1m": 10,
@@ -3135,7 +3136,7 @@
3135
  "hf_private": true
3136
  },
3137
  {
3138
- "name": "novita/sao10k/l31-70b-euryale-v2.2",
3139
  "type": "chat",
3140
  "input_price_per_1m": 1.48,
3141
  "output_price_per_1m": 1.48,
@@ -3144,7 +3145,7 @@
3144
  "size_b": 70
3145
  },
3146
  {
3147
- "name": "novita/sao10k/l3-70b-euryale-v2.1",
3148
  "type": "chat",
3149
  "input_price_per_1m": 1.48,
3150
  "output_price_per_1m": 1.48,
@@ -3165,7 +3166,7 @@
3165
  "hf_private": true
3166
  },
3167
  {
3168
- "name": "azure/openai-responses/gpt-5.2-codex",
3169
  "type": "vision",
3170
  "input_price_per_1m": 1.75,
3171
  "output_price_per_1m": 14,
@@ -3174,11 +3175,11 @@
3174
  "vision",
3175
  "reasoning"
3176
  ],
3177
- "context_window": 400000,
3178
  "hf_private": true
3179
  },
3180
  {
3181
- "name": "azure/gpt-5.3-codex",
3182
  "type": "vision",
3183
  "input_price_per_1m": 1.75,
3184
  "output_price_per_1m": 14,
@@ -3187,19 +3188,20 @@
3187
  "vision",
3188
  "reasoning"
3189
  ],
3190
- "context_window": 400000,
3191
  "hf_private": true
3192
  },
3193
  {
3194
- "name": "azure/gpt-5.2",
3195
  "type": "vision",
3196
  "input_price_per_1m": 1.75,
3197
  "output_price_per_1m": 14,
3198
  "currency": "USD",
3199
  "capabilities": [
 
3200
  "reasoning"
3201
  ],
3202
- "context_window": 200000,
3203
  "hf_private": true
3204
  },
3205
  {
@@ -3216,7 +3218,7 @@
3216
  "hf_private": true
3217
  },
3218
  {
3219
- "name": "openai-responses/gpt-5.2-codex",
3220
  "type": "vision",
3221
  "input_price_per_1m": 1.75,
3222
  "output_price_per_1m": 14,
@@ -3229,7 +3231,7 @@
3229
  "hf_private": true
3230
  },
3231
  {
3232
- "name": "openai-responses/gpt-5.2",
3233
  "type": "vision",
3234
  "input_price_per_1m": 1.75,
3235
  "output_price_per_1m": 14,
@@ -3242,7 +3244,7 @@
3242
  "hf_private": true
3243
  },
3244
  {
3245
- "name": "openai-responses/gpt-5.3-codex",
3246
  "type": "vision",
3247
  "input_price_per_1m": 1.75,
3248
  "output_price_per_1m": 14,
@@ -3255,7 +3257,7 @@
3255
  "hf_private": true
3256
  },
3257
  {
3258
- "name": "openai/gpt-5.2",
3259
  "type": "vision",
3260
  "input_price_per_1m": 1.75,
3261
  "output_price_per_1m": 14,
@@ -3268,7 +3270,7 @@
3268
  "hf_private": true
3269
  },
3270
  {
3271
- "name": "openai/gpt-5.3-chat",
3272
  "type": "vision",
3273
  "input_price_per_1m": 1.75,
3274
  "output_price_per_1m": 14,
@@ -3277,61 +3279,71 @@
3277
  "vision",
3278
  "reasoning"
3279
  ],
3280
- "context_window": 128000,
3281
  "hf_private": true
3282
  },
3283
  {
3284
- "name": "openai/gpt-5.2-chat",
3285
  "type": "vision",
3286
  "input_price_per_1m": 1.75,
3287
  "output_price_per_1m": 14,
3288
  "currency": "USD",
3289
  "capabilities": [
3290
- "vision",
3291
  "reasoning"
3292
  ],
3293
- "context_window": 128000,
3294
  "hf_private": true
3295
  },
3296
  {
3297
- "name": "azure/gpt-4.1",
 
 
 
 
 
 
 
 
 
3298
  "type": "vision",
3299
  "input_price_per_1m": 2,
3300
- "output_price_per_1m": 8,
3301
  "currency": "USD",
3302
  "capabilities": [
3303
- "vision"
 
3304
  ],
3305
- "context_window": 1047576,
3306
  "hf_private": true
3307
  },
3308
  {
3309
- "name": "azure/openai-responses/gpt-4.1",
3310
  "type": "vision",
3311
  "input_price_per_1m": 2,
3312
- "output_price_per_1m": 8,
3313
  "currency": "USD",
3314
  "capabilities": [
3315
- "vision"
 
3316
  ],
3317
- "context_window": 1047576,
3318
  "hf_private": true
3319
  },
3320
  {
3321
- "name": "perplexity/sonar-reasoning-pro",
3322
  "type": "vision",
3323
  "input_price_per_1m": 2,
3324
- "output_price_per_1m": 8,
3325
  "currency": "USD",
3326
- "context_window": 131072,
3327
  "capabilities": [
3328
  "vision",
3329
  "reasoning"
3330
  ],
 
3331
  "hf_private": true
3332
  },
3333
  {
3334
- "name": "openai-responses/gpt-4.1",
3335
  "type": "vision",
3336
  "input_price_per_1m": 2,
3337
  "output_price_per_1m": 8,
@@ -3343,7 +3355,20 @@
3343
  "hf_private": true
3344
  },
3345
  {
3346
- "name": "vertex/gemini-3-pro-preview",
 
 
 
 
 
 
 
 
 
 
 
 
 
3347
  "type": "vision",
3348
  "input_price_per_1m": 2,
3349
  "output_price_per_1m": 12,
@@ -3356,7 +3381,7 @@
3356
  "hf_private": true
3357
  },
3358
  {
3359
- "name": "vertex/gemini-3-pro-image-preview",
3360
  "type": "vision",
3361
  "input_price_per_1m": 2,
3362
  "output_price_per_1m": 12,
@@ -3369,20 +3394,16 @@
3369
  "hf_private": true
3370
  },
3371
  {
3372
- "name": "vertex/gemini-3.1-pro-preview",
3373
- "type": "vision",
3374
  "input_price_per_1m": 2,
3375
- "output_price_per_1m": 12,
3376
  "currency": "USD",
3377
- "capabilities": [
3378
- "vision",
3379
- "reasoning"
3380
- ],
3381
- "context_window": 1048576,
3382
  "hf_private": true
3383
  },
3384
  {
3385
- "name": "openai/gpt-4.1",
3386
  "type": "vision",
3387
  "input_price_per_1m": 2,
3388
  "output_price_per_1m": 8,
@@ -3394,108 +3415,86 @@
3394
  "hf_private": true
3395
  },
3396
  {
3397
- "name": "openai/o4-mini-deep-research",
3398
  "type": "vision",
3399
  "input_price_per_1m": 2,
3400
  "output_price_per_1m": 8,
3401
  "currency": "USD",
3402
  "capabilities": [
3403
- "vision",
3404
- "reasoning"
3405
  ],
3406
- "context_window": 100000,
3407
  "hf_private": true
3408
  },
3409
  {
3410
- "name": "xai/grok-4.2-beta",
3411
  "type": "vision",
3412
  "input_price_per_1m": 2,
3413
- "output_price_per_1m": 6,
3414
  "currency": "USD",
3415
  "capabilities": [
3416
- "vision",
3417
- "reasoning"
3418
  ],
3419
- "context_window": 2000000,
3420
- "hf_private": true
3421
- },
3422
- {
3423
- "name": "xai/grok-2-1212",
3424
- "type": "chat",
3425
- "input_price_per_1m": 2,
3426
- "output_price_per_1m": 10,
3427
- "currency": "USD",
3428
- "context_window": 131072,
3429
- "hf_private": true
3430
- },
3431
- {
3432
- "name": "nebius/deepseek-ai/DeepSeek-V3-0324-fast",
3433
- "type": "chat",
3434
- "input_price_per_1m": 2,
3435
- "output_price_per_1m": 6,
3436
- "currency": "USD",
3437
- "context_window": 128000,
3438
  "hf_private": true
3439
  },
3440
  {
3441
- "name": "google/gemini-3-pro-preview",
3442
  "type": "vision",
3443
  "input_price_per_1m": 2,
3444
- "output_price_per_1m": 12,
3445
  "currency": "USD",
 
3446
  "capabilities": [
3447
  "vision",
3448
  "reasoning"
3449
  ],
3450
- "context_window": 1048576,
3451
  "hf_private": true
3452
  },
3453
  {
3454
- "name": "google/gemini-3.1-pro-preview",
3455
  "type": "vision",
3456
  "input_price_per_1m": 2,
3457
- "output_price_per_1m": 12,
3458
  "currency": "USD",
3459
  "capabilities": [
3460
  "vision",
3461
  "reasoning"
3462
  ],
3463
- "context_window": 1048576,
3464
  "hf_private": true
3465
  },
3466
  {
3467
- "name": "mistral/pixtral-large-latest",
3468
  "type": "chat",
3469
  "input_price_per_1m": 2,
3470
- "output_price_per_1m": 5,
3471
  "currency": "USD",
3472
  "context_window": 131072,
3473
  "hf_private": true
3474
  },
3475
  {
3476
- "name": "azure/openai-responses/gpt-5.4",
3477
  "type": "vision",
3478
  "input_price_per_1m": 2.5,
3479
- "output_price_per_1m": 15,
3480
  "currency": "USD",
3481
  "capabilities": [
3482
- "vision",
3483
- "reasoning"
3484
  ],
3485
- "context_window": 1050000,
3486
  "hf_private": true
3487
  },
3488
  {
3489
- "name": "openai-responses/gpt-5.4",
3490
  "type": "vision",
3491
  "input_price_per_1m": 2.5,
3492
- "output_price_per_1m": 15,
3493
  "currency": "USD",
3494
  "capabilities": [
3495
- "vision",
3496
- "reasoning"
3497
  ],
3498
- "context_window": 1050000,
3499
  "hf_private": true
3500
  },
3501
  {
@@ -3536,31 +3535,33 @@
3536
  "hf_private": true
3537
  },
3538
  {
3539
- "name": "openai/gpt-4o-2024-11-20",
3540
  "type": "vision",
3541
  "input_price_per_1m": 2.5,
3542
- "output_price_per_1m": 10,
3543
  "currency": "USD",
3544
  "capabilities": [
3545
- "vision"
 
3546
  ],
3547
- "context_window": 128000,
3548
  "hf_private": true
3549
  },
3550
  {
3551
- "name": "openai/gpt-4o",
3552
  "type": "vision",
3553
  "input_price_per_1m": 2.5,
3554
- "output_price_per_1m": 10,
3555
  "currency": "USD",
3556
  "capabilities": [
3557
- "vision"
 
3558
  ],
3559
- "context_window": 128000,
3560
  "hf_private": true
3561
  },
3562
  {
3563
- "name": "anthropic/claude-sonnet-4",
3564
  "type": "vision",
3565
  "input_price_per_1m": 3,
3566
  "output_price_per_1m": 15,
@@ -3569,11 +3570,11 @@
3569
  "vision",
3570
  "reasoning"
3571
  ],
3572
- "context_window": 1000000,
3573
  "hf_private": true
3574
  },
3575
  {
3576
- "name": "anthropic/claude-sonnet-4-5",
3577
  "type": "vision",
3578
  "input_price_per_1m": 3,
3579
  "output_price_per_1m": 15,
@@ -3582,11 +3583,11 @@
3582
  "vision",
3583
  "reasoning"
3584
  ],
3585
- "context_window": 1000000,
3586
  "hf_private": true
3587
  },
3588
  {
3589
- "name": "anthropic/claude-sonnet-4-6",
3590
  "type": "vision",
3591
  "input_price_per_1m": 3,
3592
  "output_price_per_1m": 15,
@@ -3595,49 +3596,49 @@
3595
  "vision",
3596
  "reasoning"
3597
  ],
3598
- "context_window": 1000000,
3599
  "hf_private": true
3600
  },
3601
  {
3602
- "name": "together/deepseek-ai/DeepSeek-R1",
3603
- "type": "chat",
3604
  "input_price_per_1m": 3,
3605
- "output_price_per_1m": 7,
3606
  "currency": "USD",
3607
- "context_window": 64000,
3608
  "capabilities": [
3609
- "tools",
3610
- "reasoning"
3611
  ],
3612
- "hf_id": "deepseek-ai/DeepSeek-R1",
3613
- "size_b": 684.5
3614
  },
3615
  {
3616
- "name": "perplexity/sonar-pro",
3617
  "type": "vision",
3618
  "input_price_per_1m": 3,
3619
  "output_price_per_1m": 15,
3620
  "currency": "USD",
3621
- "context_window": 204800,
3622
  "capabilities": [
3623
- "vision"
 
3624
  ],
 
3625
  "hf_private": true
3626
  },
3627
  {
3628
- "name": "vertex/claude-3-5-sonnet",
3629
  "type": "vision",
3630
  "input_price_per_1m": 3,
3631
  "output_price_per_1m": 15,
3632
  "currency": "USD",
3633
  "capabilities": [
3634
- "vision"
 
3635
  ],
3636
- "context_window": 200000,
3637
  "hf_private": true
3638
  },
3639
  {
3640
- "name": "vertex/claude-sonnet-4-5",
3641
  "type": "vision",
3642
  "input_price_per_1m": 3,
3643
  "output_price_per_1m": 15,
@@ -3646,24 +3647,25 @@
3646
  "vision",
3647
  "reasoning"
3648
  ],
3649
- "context_window": 200000,
3650
  "hf_private": true
3651
  },
3652
  {
3653
- "name": "vertex/claude-3-7-sonnet",
3654
- "type": "vision",
3655
  "input_price_per_1m": 3,
3656
- "output_price_per_1m": 15,
3657
  "currency": "USD",
 
3658
  "capabilities": [
3659
- "vision",
3660
  "reasoning"
3661
  ],
3662
- "context_window": 200000,
3663
- "hf_private": true
3664
  },
3665
  {
3666
- "name": "vertex/claude-sonnet-4",
3667
  "type": "vision",
3668
  "input_price_per_1m": 3,
3669
  "output_price_per_1m": 15,
@@ -3676,19 +3678,19 @@
3676
  "hf_private": true
3677
  },
3678
  {
3679
- "name": "xai/grok-4",
3680
  "type": "vision",
3681
  "input_price_per_1m": 3,
3682
  "output_price_per_1m": 15,
3683
  "currency": "USD",
 
3684
  "capabilities": [
3685
  "vision"
3686
  ],
3687
- "context_window": 256000,
3688
  "hf_private": true
3689
  },
3690
  {
3691
- "name": "coding/claude-sonnet-4-20250514",
3692
  "type": "vision",
3693
  "input_price_per_1m": 3,
3694
  "output_price_per_1m": 15,
@@ -3714,7 +3716,7 @@
3714
  "hf_private": true
3715
  },
3716
  {
3717
- "name": "bedrock/claude-sonnet-4-5",
3718
  "type": "vision",
3719
  "input_price_per_1m": 3,
3720
  "output_price_per_1m": 15,
@@ -3723,11 +3725,11 @@
3723
  "vision",
3724
  "reasoning"
3725
  ],
3726
- "context_window": 200000,
3727
  "hf_private": true
3728
  },
3729
  {
3730
- "name": "bedrock/claude-sonnet-4-6",
3731
  "type": "vision",
3732
  "input_price_per_1m": 3,
3733
  "output_price_per_1m": 15,
@@ -3736,20 +3738,19 @@
3736
  "vision",
3737
  "reasoning"
3738
  ],
3739
- "context_window": 1000000,
3740
  "hf_private": true
3741
  },
3742
  {
3743
- "name": "bedrock/claude-sonnet-4",
3744
  "type": "vision",
3745
  "input_price_per_1m": 3,
3746
  "output_price_per_1m": 15,
3747
  "currency": "USD",
3748
  "capabilities": [
3749
- "vision",
3750
- "reasoning"
3751
  ],
3752
- "context_window": 200000,
3753
  "hf_private": true
3754
  },
3755
  {
@@ -3764,10 +3765,10 @@
3764
  "reasoning"
3765
  ],
3766
  "hf_id": "deepseek-ai/DeepSeek-R1",
3767
- "size_b": 684.5
3768
  },
3769
  {
3770
- "name": "anthropic/claude-opus-4-5",
3771
  "type": "vision",
3772
  "input_price_per_1m": 5,
3773
  "output_price_per_1m": 25,
@@ -3780,7 +3781,7 @@
3780
  "hf_private": true
3781
  },
3782
  {
3783
- "name": "anthropic/claude-opus-4-6",
3784
  "type": "vision",
3785
  "input_price_per_1m": 5,
3786
  "output_price_per_1m": 25,
@@ -3793,20 +3794,19 @@
3793
  "hf_private": true
3794
  },
3795
  {
3796
- "name": "vertex/claude-opus-4-6",
3797
  "type": "vision",
3798
  "input_price_per_1m": 5,
3799
- "output_price_per_1m": 25,
3800
  "currency": "USD",
3801
  "capabilities": [
3802
- "vision",
3803
- "reasoning"
3804
  ],
3805
- "context_window": 1000000,
3806
  "hf_private": true
3807
  },
3808
  {
3809
- "name": "vertex/claude-opus-4-5",
3810
  "type": "vision",
3811
  "input_price_per_1m": 5,
3812
  "output_price_per_1m": 25,
@@ -3815,31 +3815,20 @@
3815
  "vision",
3816
  "reasoning"
3817
  ],
3818
- "context_window": 200000,
3819
  "hf_private": true
3820
  },
3821
  {
3822
- "name": "openai/chatgpt-4o",
3823
  "type": "vision",
3824
  "input_price_per_1m": 5,
3825
- "output_price_per_1m": 15,
3826
- "currency": "USD",
3827
- "capabilities": [
3828
- "vision"
3829
- ],
3830
- "context_window": 128000,
3831
- "hf_private": true
3832
- },
3833
- {
3834
- "name": "xai/grok-3",
3835
- "type": "chat",
3836
- "input_price_per_1m": 5,
3837
  "output_price_per_1m": 25,
3838
  "currency": "USD",
3839
- "context_window": 131072,
3840
  "capabilities": [
3841
- "tools"
 
3842
  ],
 
3843
  "hf_private": true
3844
  },
3845
  {
@@ -3868,6 +3857,18 @@
3868
  "context_window": 1000000,
3869
  "hf_private": true
3870
  },
 
 
 
 
 
 
 
 
 
 
 
 
3871
  {
3872
  "name": "openai/o3-deep-research",
3873
  "type": "vision",
@@ -3882,7 +3883,7 @@
3882
  "hf_private": true
3883
  },
3884
  {
3885
- "name": "anthropic/claude-opus-4-1",
3886
  "type": "vision",
3887
  "input_price_per_1m": 15,
3888
  "output_price_per_1m": 75,
@@ -3895,7 +3896,7 @@
3895
  "hf_private": true
3896
  },
3897
  {
3898
- "name": "anthropic/claude-opus-4",
3899
  "type": "vision",
3900
  "input_price_per_1m": 15,
3901
  "output_price_per_1m": 75,
@@ -3908,20 +3909,20 @@
3908
  "hf_private": true
3909
  },
3910
  {
3911
- "name": "openai-responses/gpt-5-pro",
3912
  "type": "vision",
3913
  "input_price_per_1m": 15,
3914
- "output_price_per_1m": 120,
3915
  "currency": "USD",
3916
  "capabilities": [
3917
  "vision",
3918
  "reasoning"
3919
  ],
3920
- "context_window": 400000,
3921
  "hf_private": true
3922
  },
3923
  {
3924
- "name": "vertex/claude-opus-4-1",
3925
  "type": "vision",
3926
  "input_price_per_1m": 15,
3927
  "output_price_per_1m": 75,
@@ -3934,7 +3935,7 @@
3934
  "hf_private": true
3935
  },
3936
  {
3937
- "name": "vertex/claude-opus-4",
3938
  "type": "vision",
3939
  "input_price_per_1m": 15,
3940
  "output_price_per_1m": 75,
@@ -3947,16 +3948,16 @@
3947
  "hf_private": true
3948
  },
3949
  {
3950
- "name": "openai/o1",
3951
  "type": "vision",
3952
  "input_price_per_1m": 15,
3953
- "output_price_per_1m": 60,
3954
  "currency": "USD",
3955
  "capabilities": [
3956
  "vision",
3957
  "reasoning"
3958
  ],
3959
- "context_window": 200000,
3960
  "hf_private": true
3961
  },
3962
  {
@@ -3999,7 +4000,7 @@
3999
  "hf_private": true
4000
  },
4001
  {
4002
- "name": "azure/openai-responses/gpt-5.4-pro",
4003
  "type": "vision",
4004
  "input_price_per_1m": 30,
4005
  "output_price_per_1m": 180,
@@ -4025,7 +4026,7 @@
4025
  "hf_private": true
4026
  },
4027
  {
4028
- "name": "openai/gpt-5.4-pro",
4029
  "type": "vision",
4030
  "input_price_per_1m": 30,
4031
  "output_price_per_1m": 180,
@@ -4347,7 +4348,9 @@
4347
  "input_price_per_1m": 0.75,
4348
  "output_price_per_1m": 2.25,
4349
  "currency": "USD",
4350
- "flavor": "fast"
 
 
4351
  },
4352
  {
4353
  "name": "DeepSeek-V3-0324 (base)",
@@ -5562,7 +5565,8 @@
5562
  "input_price_per_1m": 0.065,
5563
  "output_price_per_1m": 0.14,
5564
  "currency": "USD",
5565
- "hf_id": "microsoft/phi-4"
 
5566
  },
5567
  {
5568
  "name": "baidu/ernie-4.5-21b-a3b-thinking",
@@ -7148,7 +7152,8 @@
7148
  "hf_id": "deepseek-ai/DeepSeek-V3",
7149
  "capabilities": [
7150
  "tools"
7151
- ]
 
7152
  },
7153
  {
7154
  "name": "qwen/qwen3.5-397b-a17b",
 
998
  "eu_endpoints": true,
999
  "models": [
1000
  {
1001
+ "name": "novita/meta-llama/llama-3.2-1b-instruct",
1002
  "type": "chat",
1003
  "input_price_per_1m": 0.02,
1004
+ "output_price_per_1m": 0.02,
1005
  "currency": "USD",
1006
+ "context_window": 131000,
1007
+ "size_b": 1
1008
  },
1009
  {
1010
+ "name": "deepinfra/meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
1011
  "type": "chat",
1012
  "input_price_per_1m": 0.02,
1013
+ "output_price_per_1m": 0.05,
1014
  "currency": "USD",
1015
+ "context_window": 131072,
1016
+ "size_b": 8
1017
  },
1018
  {
1019
  "name": "openai/gpt-5-nano",
 
1051
  "hf_id": "meta-llama/Meta-Llama-3-8B-Instruct"
1052
  },
1053
  {
1054
+ "name": "openai-responses/gpt-5-nano",
1055
  "type": "vision",
1056
  "input_price_per_1m": 0.05,
1057
  "output_price_per_1m": 0.4,
1058
  "currency": "USD",
1059
  "capabilities": [
1060
+ "vision",
1061
  "reasoning"
1062
  ],
1063
+ "context_window": 400000,
 
 
 
 
 
 
 
 
 
 
 
 
1064
  "hf_private": true
1065
  },
1066
  {
1067
+ "name": "azure/gpt-5-nano",
1068
  "type": "vision",
1069
  "input_price_per_1m": 0.05,
1070
  "output_price_per_1m": 0.4,
1071
  "currency": "USD",
1072
  "capabilities": [
 
1073
  "reasoning"
1074
  ],
1075
+ "context_window": 200000,
1076
  "hf_private": true
1077
  },
1078
  {
1079
+ "name": "novita/sao10k/l3-8b-lunaris",
1080
  "type": "chat",
1081
  "input_price_per_1m": 0.05,
1082
  "output_price_per_1m": 0.05,
 
1098
  "hf_id": "meta-llama/Meta-Llama-3.1-8B-Instruct"
1099
  },
1100
  {
1101
+ "name": "novita/Sao10K/L3-8B-Stheno-v3.2",
1102
  "type": "chat",
1103
  "input_price_per_1m": 0.05,
1104
  "output_price_per_1m": 0.05,
 
1106
  "context_window": 8192,
1107
  "size_b": 8
1108
  },
1109
+ {
1110
+ "name": "alibaba/qwen-turbo",
1111
+ "type": "chat",
1112
+ "input_price_per_1m": 0.05,
1113
+ "output_price_per_1m": 0.2,
1114
+ "currency": "USD",
1115
+ "context_window": 1000000,
1116
+ "capabilities": [
1117
+ "tools"
1118
+ ],
1119
+ "hf_private": true
1120
+ },
1121
  {
1122
  "name": "together/meta-llama/Llama-3.2-3B-Instruct-Turbo",
1123
  "type": "chat",
 
1128
  "size_b": 3
1129
  },
1130
  {
1131
+ "name": "deepinfra/microsoft/phi-4",
1132
  "type": "chat",
1133
  "input_price_per_1m": 0.07,
1134
+ "output_price_per_1m": 0.14,
1135
  "currency": "USD",
1136
  "context_window": 16384,
1137
+ "hf_private": true
1138
  },
1139
  {
1140
+ "name": "deepinfra/Qwen/Qwen2.5-Coder-32B-Instruct",
1141
  "type": "chat",
1142
  "input_price_per_1m": 0.07,
1143
+ "output_price_per_1m": 0.16,
1144
  "currency": "USD",
1145
  "context_window": 16384,
1146
+ "size_b": 32
1147
  },
1148
  {
1149
  "name": "novita/gryphe/mythomax-l2-13b",
 
1155
  "size_b": 13
1156
  },
1157
  {
1158
+ "name": "mistral/devstral-small-2507",
1159
+ "type": "chat",
 
 
 
 
 
 
 
 
 
 
 
 
1160
  "input_price_per_1m": 0.1,
1161
+ "output_price_per_1m": 0.3,
1162
  "currency": "USD",
1163
+ "context_window": 131072,
1164
  "capabilities": [
1165
+ "tools"
1166
  ],
1167
+ "hf_id": "mistralai/Devstral-Small-2507",
1168
+ "size_b": 24
 
 
 
 
 
 
 
 
 
1169
  },
1170
  {
1171
+ "name": "mistral/devstral-small-latest",
1172
  "type": "chat",
1173
  "input_price_per_1m": 0.1,
1174
  "output_price_per_1m": 0.3,
1175
  "currency": "USD",
1176
+ "context_window": 131072,
 
1177
  "capabilities": [
1178
+ "tools"
 
1179
  ],
1180
+ "hf_id": "mistralai/Devstral-Small-2507",
1181
+ "size_b": 24
1182
  },
1183
  {
1184
+ "name": "mistral/mistral-small-2503",
1185
+ "type": "chat",
1186
  "input_price_per_1m": 0.1,
1187
+ "output_price_per_1m": 0.3,
1188
  "currency": "USD",
1189
+ "context_window": 32768,
 
 
 
1190
  "hf_private": true
1191
  },
1192
  {
 
1240
  "hf_private": true
1241
  },
1242
  {
1243
+ "name": "together/meta-llama/Meta-Llama-3-8B-Instruct-Lite",
1244
  "type": "chat",
1245
  "input_price_per_1m": 0.1,
1246
+ "output_price_per_1m": 0.1,
1247
+ "currency": "USD",
1248
+ "context_window": 8192,
1249
+ "size_b": 8
1250
+ },
1251
+ {
1252
+ "name": "groq/openai/gpt-oss-20b",
1253
+ "type": "chat",
1254
+ "input_price_per_1m": 0.1,
1255
+ "output_price_per_1m": 0.5,
1256
  "currency": "USD",
1257
  "context_window": 131072,
1258
+ "size_b": 20,
1259
  "capabilities": [
1260
+ "tools",
1261
+ "reasoning"
1262
  ],
1263
+ "hf_id": "openai/gpt-oss-20b"
 
1264
  },
1265
  {
1266
+ "name": "openai-responses/gpt-4.1-nano",
1267
+ "type": "vision",
1268
  "input_price_per_1m": 0.1,
1269
+ "output_price_per_1m": 0.4,
1270
  "currency": "USD",
1271
+ "capabilities": [
1272
+ "vision"
1273
+ ],
1274
+ "context_window": 1047576,
1275
  "hf_private": true
1276
  },
1277
  {
1278
+ "name": "azure/gpt-4.1-nano",
1279
+ "type": "vision",
1280
  "input_price_per_1m": 0.1,
1281
+ "output_price_per_1m": 0.4,
1282
  "currency": "USD",
 
1283
  "capabilities": [
1284
+ "vision"
1285
  ],
1286
+ "context_window": 1047576,
1287
+ "hf_private": true
1288
  },
1289
  {
1290
+ "name": "azure/openai-responses/gpt-4.1-nano",
1291
+ "type": "vision",
1292
+ "input_price_per_1m": 0.1,
1293
+ "output_price_per_1m": 0.4,
1294
+ "currency": "USD",
1295
+ "capabilities": [
1296
+ "vision"
1297
+ ],
1298
+ "context_window": 1047576,
1299
+ "hf_private": true
1300
+ },
1301
+ {
1302
+ "name": "deepinfra/Qwen/Qwen3-32B",
1303
  "type": "chat",
1304
  "input_price_per_1m": 0.1,
1305
+ "output_price_per_1m": 0.3,
1306
  "currency": "USD",
1307
+ "context_window": 40000,
1308
+ "size_b": 32,
1309
  "capabilities": [
1310
  "tools",
1311
  "reasoning"
1312
  ],
1313
+ "hf_id": "Qwen/Qwen3-32B"
1314
  },
1315
  {
1316
  "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct-Turbo",
 
1361
  "size_b": 8
1362
  },
1363
  {
1364
+ "name": "mistral/mistral-small-latest",
 
 
 
 
 
 
 
 
 
1365
  "type": "vision",
1366
  "input_price_per_1m": 0.15,
1367
  "output_price_per_1m": 0.6,
1368
  "currency": "USD",
1369
  "capabilities": [
1370
+ "vision",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1371
  "reasoning"
1372
  ],
1373
+ "context_window": 256000,
1374
+ "hf_private": true
1375
  },
1376
  {
1377
+ "name": "mistral/mistral-small-2603",
1378
+ "type": "vision",
1379
  "input_price_per_1m": 0.15,
1380
  "output_price_per_1m": 0.6,
1381
  "currency": "USD",
1382
  "capabilities": [
1383
+ "vision",
1384
  "reasoning"
1385
  ],
1386
+ "context_window": 256000,
1387
+ "size_b": 119,
1388
+ "hf_id": "mistralai/Mistral-Small-4-119B-2603"
1389
  },
1390
  {
1391
+ "name": "openai/gpt-4o-mini",
1392
  "type": "vision",
1393
  "input_price_per_1m": 0.15,
1394
  "output_price_per_1m": 0.6,
1395
  "currency": "USD",
1396
  "capabilities": [
1397
+ "vision"
 
1398
  ],
1399
+ "context_window": 128000,
1400
  "hf_private": true
1401
  },
1402
  {
1403
+ "name": "nebius/openai/gpt-oss-120b",
1404
+ "type": "chat",
1405
  "input_price_per_1m": 0.15,
1406
  "output_price_per_1m": 0.6,
1407
  "currency": "USD",
1408
  "capabilities": [
 
1409
  "reasoning"
1410
  ],
1411
+ "context_window": 131000,
1412
+ "size_b": 120,
1413
+ "hf_id": "openai/gpt-oss-120b"
1414
  },
1415
  {
1416
  "name": "groq/openai/gpt-oss-120b",
 
1427
  "hf_id": "openai/gpt-oss-120b"
1428
  },
1429
  {
1430
+ "name": "novita/deepseek/deepseek-r1-distill-qwen-14b",
1431
  "type": "chat",
1432
+ "input_price_per_1m": 0.15,
1433
+ "output_price_per_1m": 0.15,
1434
  "currency": "USD",
1435
+ "context_window": 128000,
1436
+ "size_b": 671,
1437
  "capabilities": [
1438
+ "tools",
1439
+ "reasoning"
1440
  ],
1441
+ "hf_id": "deepseek-ai/DeepSeek-R1"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1442
  },
1443
  {
1444
+ "name": "parasail/parasail-qwen3-235b-a22b-instruct-2507",
1445
  "type": "chat",
1446
+ "input_price_per_1m": 0.15,
1447
+ "output_price_per_1m": 0.85,
1448
  "currency": "USD",
1449
+ "context_window": 262144,
1450
+ "size_b": 235
 
 
 
 
 
1451
  },
1452
  {
1453
+ "name": "novita/mistralai/mistral-nemo",
1454
  "type": "chat",
1455
+ "input_price_per_1m": 0.17,
1456
+ "output_price_per_1m": 0.17,
1457
  "currency": "USD",
1458
  "context_window": 131072,
1459
  "capabilities": [
1460
+ "tools"
 
1461
  ],
1462
+ "hf_id": "mistralai/Mistral-Nemo-Instruct-2407",
1463
+ "size_b": 12
1464
  },
1465
  {
1466
+ "name": "together/meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
1467
+ "type": "chat",
1468
+ "input_price_per_1m": 0.18,
1469
+ "output_price_per_1m": 0.18,
1470
  "currency": "USD",
 
 
 
1471
  "context_window": 131072,
1472
+ "size_b": 8
 
1473
  },
1474
  {
1475
+ "name": "openai/gpt-5.4-nano",
1476
  "type": "vision",
1477
  "input_price_per_1m": 0.2,
1478
  "output_price_per_1m": 1.25,
 
1485
  "hf_private": true
1486
  },
1487
  {
1488
+ "name": "together/meta-llama/LlamaGuard-2-8b",
1489
+ "type": "chat",
1490
+ "input_price_per_1m": 0.2,
1491
+ "output_price_per_1m": 0.2,
1492
+ "currency": "USD",
1493
+ "context_window": 8192,
1494
+ "size_b": 8
1495
+ },
1496
+ {
1497
+ "name": "openai-responses/gpt-5.4-nano",
1498
  "type": "vision",
1499
  "input_price_per_1m": 0.2,
1500
  "output_price_per_1m": 1.25,
 
1535
  "hf_id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
1536
  },
1537
  {
1538
+ "name": "deepinfra/Qwen/Qwen3-235B-A22B",
1539
+ "type": "chat",
1540
+ "input_price_per_1m": 0.2,
1541
+ "output_price_per_1m": 0.6,
1542
+ "currency": "USD",
1543
+ "context_window": 40000,
1544
+ "size_b": 235,
1545
+ "capabilities": [
1546
+ "tools",
1547
+ "reasoning"
1548
+ ],
1549
+ "hf_id": "Qwen/Qwen3-235B-A22B"
1550
+ },
1551
+ {
1552
+ "name": "deepinfra/zai-org/GLM-4.5-Air",
1553
+ "type": "chat",
1554
+ "input_price_per_1m": 0.2,
1555
+ "output_price_per_1m": 1.1,
1556
+ "currency": "USD",
1557
+ "context_window": 131072,
1558
+ "capabilities": [
1559
+ "tools",
1560
+ "reasoning"
1561
+ ],
1562
+ "hf_id": "zai-org/GLM-4.5-Air",
1563
+ "size_b": 110.5
1564
+ },
1565
+ {
1566
+ "name": "xai/grok-4-1-fast-reasoning",
1567
  "type": "vision",
1568
  "input_price_per_1m": 0.2,
1569
  "output_price_per_1m": 0.5,
 
1575
  "hf_private": true
1576
  },
1577
  {
1578
+ "name": "xai/grok-4-1-fast-non-reasoning",
1579
  "type": "vision",
1580
  "input_price_per_1m": 0.2,
1581
  "output_price_per_1m": 0.5,
1582
  "currency": "USD",
 
1583
  "capabilities": [
1584
+ "vision"
 
 
1585
  ],
1586
+ "context_window": 2000000,
1587
  "hf_private": true
1588
  },
1589
  {
1590
+ "name": "xai/grok-code-fast-1",
1591
  "type": "vision",
1592
  "input_price_per_1m": 0.2,
1593
+ "output_price_per_1m": 1.5,
1594
  "currency": "USD",
1595
  "capabilities": [
1596
  "vision"
1597
  ],
1598
+ "context_window": 256000,
1599
+ "hf_private": true
1600
+ },
1601
+ {
1602
+ "name": "xai/grok-4-fast",
1603
+ "type": "vision",
1604
+ "input_price_per_1m": 0.2,
1605
+ "output_price_per_1m": 0.5,
1606
+ "currency": "USD",
1607
  "context_window": 2000000,
1608
+ "capabilities": [
1609
+ "vision",
1610
+ "tools",
1611
+ "reasoning"
1612
+ ],
1613
  "hf_private": true
1614
  },
1615
  {
 
1627
  "hf_private": true
1628
  },
1629
  {
1630
+ "name": "alibaba/qwen3-30b-a3b-instruct-2507",
1631
  "type": "vision",
1632
  "input_price_per_1m": 0.2,
1633
+ "output_price_per_1m": 0.8,
1634
  "currency": "USD",
1635
  "capabilities": [
1636
  "vision"
1637
  ],
1638
+ "context_window": 131072,
1639
+ "size_b": 30,
1640
+ "hf_id": "Qwen/Qwen3-30B-A3B-Instruct-2507"
1641
  },
1642
  {
1643
+ "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct",
1644
+ "type": "chat",
1645
  "input_price_per_1m": 0.23,
1646
  "output_price_per_1m": 0.4,
1647
  "currency": "USD",
1648
  "context_window": 131072,
1649
+ "size_b": 70,
1650
  "capabilities": [
1651
+ "tools"
1652
  ],
1653
+ "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
1654
  },
1655
  {
1656
  "name": "deepinfra/deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
 
1666
  "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
1667
  },
1668
  {
1669
+ "name": "deepinfra/Qwen/Qwen2.5-72B-Instruct",
1670
+ "type": "vision",
1671
  "input_price_per_1m": 0.23,
1672
  "output_price_per_1m": 0.4,
1673
  "currency": "USD",
1674
+ "context_window": 131072,
1675
+ "size_b": 72,
1676
+ "capabilities": [
1677
+ "vision"
1678
+ ],
1679
+ "hf_id": "Qwen/Qwen2.5-VL-72B-Instruct"
1680
  },
1681
  {
1682
+ "name": "deepinfra/meta-llama/Meta-Llama-3.1-70B-Instruct",
1683
  "type": "chat",
1684
  "input_price_per_1m": 0.23,
1685
  "output_price_per_1m": 0.4,
1686
  "currency": "USD",
1687
+ "context_window": 130815,
1688
+ "size_b": 70
 
 
 
 
1689
  },
1690
  {
1691
+ "name": "mistral/open-mistral-7b",
1692
+ "type": "chat",
1693
  "input_price_per_1m": 0.25,
1694
+ "output_price_per_1m": 0.25,
1695
  "currency": "USD",
1696
+ "context_window": 32768,
1697
+ "size_b": 7
 
 
 
1698
  },
1699
  {
1700
+ "name": "vertex/gemini-3.1-flash-lite-preview",
1701
  "type": "vision",
1702
  "input_price_per_1m": 0.25,
1703
+ "output_price_per_1m": 1.5,
1704
  "currency": "USD",
1705
  "capabilities": [
1706
+ "vision"
 
1707
  ],
1708
+ "context_window": 1048576,
1709
  "hf_private": true
1710
  },
1711
  {
1712
+ "name": "google/gemini-3.1-flash-lite-preview",
1713
  "type": "vision",
1714
  "input_price_per_1m": 0.25,
1715
  "output_price_per_1m": 1.5,
 
1721
  "hf_private": true
1722
  },
1723
  {
1724
+ "name": "openai-responses/gpt-5-mini",
1725
  "type": "vision",
1726
  "input_price_per_1m": 0.25,
1727
+ "output_price_per_1m": 2,
1728
  "currency": "USD",
1729
  "capabilities": [
1730
+ "vision",
1731
+ "reasoning"
1732
  ],
1733
+ "context_window": 400000,
1734
  "hf_private": true
1735
  },
1736
  {
1737
+ "name": "azure/gpt-5-mini",
1738
+ "type": "vision",
1739
  "input_price_per_1m": 0.25,
1740
+ "output_price_per_1m": 2,
1741
  "currency": "USD",
1742
+ "capabilities": [
1743
+ "reasoning"
1744
+ ],
1745
+ "context_window": 200000,
1746
+ "hf_private": true
1747
  },
1748
  {
1749
  "name": "novita/deepseek/deepseek-v3.2",
 
1768
  "capabilities": [
1769
  "tools"
1770
  ],
1771
+ "hf_id": "deepseek-ai/DeepSeek-V3",
1772
+ "size_b": 671
1773
  },
1774
  {
1775
  "name": "deepseek/deepseek-reasoner",
 
1781
  "hf_private": true
1782
  },
1783
  {
1784
+ "name": "mistral/codestral-latest",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1785
  "type": "chat",
1786
  "input_price_per_1m": 0.3,
1787
+ "output_price_per_1m": 0.9,
 
 
 
 
 
 
 
 
 
1788
  "currency": "USD",
1789
+ "context_window": 131072,
 
 
 
1790
  "hf_private": true
1791
  },
1792
  {
 
1816
  "hf_private": true
1817
  },
1818
  {
1819
+ "name": "minimaxi/MiniMax-M2.7",
1820
  "type": "vision",
1821
  "input_price_per_1m": 0.3,
1822
  "output_price_per_1m": 1.2,
 
1830
  "hf_private": true
1831
  },
1832
  {
1833
+ "name": "minimaxi/MiniMax-M2",
1834
  "type": "chat",
1835
  "input_price_per_1m": 0.3,
1836
+ "output_price_per_1m": 1.2,
1837
  "currency": "USD",
1838
+ "context_window": 200000,
 
1839
  "capabilities": [
1840
+ "tools",
1841
  "reasoning"
1842
  ],
1843
+ "hf_id": "MiniMaxAI/MiniMax-M2",
1844
+ "size_b": 228.7
1845
  },
1846
  {
1847
+ "name": "minimaxi/MiniMax-M2.5",
1848
  "type": "vision",
1849
  "input_price_per_1m": 0.3,
1850
  "output_price_per_1m": 1.2,
 
1854
  "reasoning"
1855
  ],
1856
  "context_window": 200000,
1857
+ "hf_id": "MiniMaxAI/MiniMax-M2.5",
1858
+ "size_b": 228.7
 
 
 
 
 
 
 
 
 
 
 
 
1859
  },
1860
  {
1861
+ "name": "google/gemini-2.5-flash",
1862
  "type": "vision",
1863
  "input_price_per_1m": 0.3,
1864
  "output_price_per_1m": 2.5,
 
1897
  "size_b": 685.4
1898
  },
1899
  {
1900
+ "name": "together/Qwen/Qwen2.5-7B-Instruct-Turbo",
1901
+ "type": "chat",
1902
+ "input_price_per_1m": 0.3,
1903
+ "output_price_per_1m": 0.3,
1904
+ "currency": "USD",
1905
+ "context_window": 32768,
1906
+ "size_b": 7
1907
+ },
1908
+ {
1909
+ "name": "novita/minimax-m2.7",
1910
  "type": "vision",
1911
  "input_price_per_1m": 0.3,
1912
  "output_price_per_1m": 1.2,
 
1920
  "hf_private": true
1921
  },
1922
  {
1923
+ "name": "novita/MiniMax-M2.7",
1924
+ "type": "vision",
1925
  "input_price_per_1m": 0.3,
1926
  "output_price_per_1m": 1.2,
1927
  "currency": "USD",
1928
+ "capabilities": [
1929
+ "vision",
1930
+ "reasoning"
1931
+ ],
1932
  "context_window": 200000,
1933
+ "hf_id": "MiniMaxAI/MiniMax-M2.7",
1934
+ "hf_private": true
1935
+ },
1936
+ {
1937
+ "name": "novita/deepseek/deepseek-r1-distill-qwen-32b",
1938
+ "type": "chat",
1939
+ "input_price_per_1m": 0.3,
1940
+ "output_price_per_1m": 0.3,
1941
+ "currency": "USD",
1942
+ "context_window": 12800,
1943
+ "size_b": 32,
1944
+ "capabilities": [
1945
+ "reasoning"
1946
+ ],
1947
+ "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
1948
+ },
1949
+ {
1950
+ "name": "coding/gemini-2.5-flash",
1951
+ "type": "vision",
1952
+ "input_price_per_1m": 0.3,
1953
+ "output_price_per_1m": 2.5,
1954
+ "currency": "USD",
1955
+ "capabilities": [
1956
+ "vision",
1957
+ "reasoning"
1958
+ ],
1959
+ "context_window": 1048576,
1960
+ "hf_private": true
1961
+ },
1962
+ {
1963
+ "name": "deepinfra/deepseek-ai/DeepSeek-V3.1",
1964
+ "type": "chat",
1965
+ "input_price_per_1m": 0.3,
1966
+ "output_price_per_1m": 1,
1967
+ "currency": "USD",
1968
+ "context_window": 163840,
1969
  "capabilities": [
1970
  "tools",
1971
  "reasoning"
1972
  ],
1973
+ "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
1974
+ "size_b": 671
1975
+ },
1976
+ {
1977
+ "name": "parasail/parasail-gemma3-27b-it",
1978
+ "type": "chat",
1979
+ "input_price_per_1m": 0.3,
1980
+ "output_price_per_1m": 0.5,
1981
+ "currency": "USD",
1982
+ "context_window": 128000,
1983
+ "size_b": 27
1984
  },
1985
  {
1986
+ "name": "xai/grok-3-mini",
1987
+ "type": "chat",
1988
  "input_price_per_1m": 0.3,
1989
+ "output_price_per_1m": 0.5,
1990
  "currency": "USD",
1991
  "capabilities": [
 
1992
  "reasoning"
1993
  ],
1994
+ "context_window": 131072,
1995
+ "hf_private": true
 
1996
  },
1997
  {
1998
+ "name": "alibaba/qwen3-coder-flash",
1999
  "type": "vision",
2000
  "input_price_per_1m": 0.3,
2001
+ "output_price_per_1m": 1.5,
2002
  "currency": "USD",
2003
  "capabilities": [
2004
+ "vision"
 
2005
  ],
2006
  "context_window": 1048576,
2007
  "hf_private": true
2008
  },
 
 
 
 
 
 
 
 
 
2009
  {
2010
  "name": "deepinfra/meta-llama/Llama-3.2-90B-Vision-Instruct",
2011
  "type": "chat",
 
2042
  "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2043
  },
2044
  {
2045
+ "name": "mistral/devstral-latest",
2046
+ "type": "chat",
2047
  "input_price_per_1m": 0.4,
2048
+ "output_price_per_1m": 2,
2049
+ "currency": "USD",
2050
+ "context_window": 256000,
2051
+ "hf_private": true
2052
+ },
2053
+ {
2054
+ "name": "mistral/mistral-medium-latest",
2055
+ "type": "chat",
2056
+ "input_price_per_1m": 0.4,
2057
+ "output_price_per_1m": 2,
2058
+ "currency": "USD",
2059
+ "context_window": 131072,
2060
+ "hf_private": true
2061
+ },
2062
+ {
2063
+ "name": "mistral/devstral-medium-2507",
2064
+ "type": "chat",
2065
+ "input_price_per_1m": 0.4,
2066
+ "output_price_per_1m": 2,
2067
  "currency": "USD",
2068
+ "context_window": 131072,
2069
  "capabilities": [
2070
+ "tools"
2071
  ],
 
2072
  "hf_private": true
2073
  },
2074
  {
2075
+ "name": "openai/gpt-4.1-mini",
2076
  "type": "vision",
2077
  "input_price_per_1m": 0.4,
2078
  "output_price_per_1m": 1.6,
 
2084
  "hf_private": true
2085
  },
2086
  {
2087
+ "name": "nebius/Qwen/Qwen3-Coder-480B-A35B-Instruct",
2088
  "type": "chat",
2089
  "input_price_per_1m": 0.4,
2090
+ "output_price_per_1m": 1.8,
2091
  "currency": "USD",
2092
+ "context_window": 262000,
2093
  "size_b": 480,
2094
  "capabilities": [
2095
  "tools"
 
2097
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2098
  },
2099
  {
2100
+ "name": "nebius/zai-org/GLM-4.7",
2101
  "type": "chat",
2102
  "input_price_per_1m": 0.4,
2103
+ "output_price_per_1m": 2,
2104
  "currency": "USD",
 
2105
  "capabilities": [
2106
+ "reasoning"
2107
  ],
2108
+ "context_window": 200000,
2109
+ "hf_id": "zai-org/GLM-4.7",
2110
+ "size_b": 358.3
2111
  },
2112
  {
2113
  "name": "openai-responses/gpt-4.1-mini",
 
2122
  "hf_private": true
2123
  },
2124
  {
2125
+ "name": "azure/gpt-4.1-mini",
2126
+ "type": "vision",
2127
+ "input_price_per_1m": 0.4,
2128
+ "output_price_per_1m": 1.6,
2129
+ "currency": "USD",
2130
+ "capabilities": [
2131
+ "vision"
2132
+ ],
2133
+ "context_window": 1047576,
2134
+ "hf_private": true
2135
+ },
2136
+ {
2137
+ "name": "azure/openai-responses/gpt-4.1-mini",
2138
  "type": "vision",
2139
  "input_price_per_1m": 0.4,
2140
  "output_price_per_1m": 1.6,
 
2169
  "hf_private": true
2170
  },
2171
  {
2172
+ "name": "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct",
 
 
 
 
 
 
 
 
 
 
 
 
 
2173
  "type": "chat",
2174
  "input_price_per_1m": 0.4,
2175
+ "output_price_per_1m": 1.6,
2176
  "currency": "USD",
2177
+ "context_window": 262144,
2178
  "size_b": 480,
2179
  "capabilities": [
2180
  "tools"
 
2182
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2183
  },
2184
  {
2185
+ "name": "alibaba/qwen-plus",
2186
  "type": "chat",
2187
  "input_price_per_1m": 0.4,
2188
+ "output_price_per_1m": 1.2,
2189
  "currency": "USD",
2190
  "context_window": 131072,
2191
+ "capabilities": [
2192
+ "tools"
2193
+ ],
2194
  "hf_private": true
2195
  },
2196
  {
2197
+ "name": "mistral/mistral-large-latest",
2198
  "type": "chat",
2199
+ "input_price_per_1m": 0.5,
2200
+ "output_price_per_1m": 1.5,
2201
  "currency": "USD",
2202
  "context_window": 131072,
2203
  "capabilities": [
 
2205
  ],
2206
  "hf_private": true
2207
  },
 
 
 
 
 
 
 
 
 
2208
  {
2209
  "name": "vertex/gemini-3.1-flash-image-preview",
2210
  "type": "vision",
 
2231
  "context_window": 1048576,
2232
  "hf_private": true
2233
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2234
  {
2235
  "name": "google/gemini-3.1-flash-image-preview",
2236
  "type": "vision",
 
2258
  "hf_private": true
2259
  },
2260
  {
2261
+ "name": "nebius/moonshotai/kimi-k2.5",
2262
+ "type": "vision",
2263
  "input_price_per_1m": 0.5,
2264
+ "output_price_per_1m": 2.5,
2265
  "currency": "USD",
 
2266
  "capabilities": [
2267
+ "vision",
2268
+ "reasoning"
2269
  ],
2270
+ "context_window": 256000,
2271
+ "hf_id": "moonshotai/Kimi-K2.5",
2272
+ "size_b": 1058.6
2273
  },
2274
  {
2275
+ "name": "nebius/moonshotai/Kimi-K2-Instruct",
2276
  "type": "chat",
2277
+ "input_price_per_1m": 0.5,
2278
+ "output_price_per_1m": 2.4,
2279
  "currency": "USD",
2280
+ "context_window": 131000,
 
2281
  "capabilities": [
2282
  "tools"
2283
  ],
2284
+ "hf_id": "moonshotai/Kimi-K2-Instruct",
2285
+ "size_b": 1026.5
 
 
 
 
 
 
 
 
 
 
 
2286
  },
2287
  {
2288
+ "name": "nebius/deepseek-ai/DeepSeek-V3-0324",
2289
+ "type": "chat",
2290
+ "input_price_per_1m": 0.5,
2291
+ "output_price_per_1m": 1.5,
2292
  "currency": "USD",
2293
+ "context_window": 128000,
2294
  "capabilities": [
2295
+ "tools",
2296
  "reasoning"
2297
  ],
2298
+ "size_b": 685,
2299
+ "hf_id": "deepseek-ai/DeepSeek-V3-0324"
 
2300
  },
2301
  {
2302
+ "name": "novita/meta-llama/llama-3-70b-instruct",
2303
  "type": "chat",
2304
+ "input_price_per_1m": 0.51,
2305
+ "output_price_per_1m": 0.74,
2306
  "currency": "USD",
2307
+ "context_window": 8192,
2308
+ "size_b": 70,
2309
  "capabilities": [
2310
  "tools"
2311
  ],
2312
+ "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2313
  },
2314
  {
2315
+ "name": "openai/o4-mini",
2316
  "type": "vision",
2317
+ "input_price_per_1m": 0.55,
2318
+ "output_price_per_1m": 2.2,
2319
  "currency": "USD",
2320
  "capabilities": [
 
2321
  "reasoning"
2322
  ],
2323
+ "context_window": 200000,
2324
+ "hf_private": true
 
2325
  },
2326
  {
2327
+ "name": "vertex/deepseek-v3.2",
2328
+ "type": "vision",
2329
+ "input_price_per_1m": 0.56,
2330
+ "output_price_per_1m": 1.68,
2331
  "currency": "USD",
 
2332
  "capabilities": [
2333
+ "vision",
2334
  "reasoning"
2335
  ],
2336
+ "context_window": 163840,
2337
+ "hf_id": "deepseek-ai/DeepSeek-V3.2",
2338
+ "size_b": 685.4
2339
  },
2340
  {
2341
+ "name": "novita/moonshotai/kimi-k2-instruct",
2342
  "type": "chat",
2343
+ "input_price_per_1m": 0.57,
2344
+ "output_price_per_1m": 2.3,
2345
  "currency": "USD",
2346
+ "context_window": 131072,
2347
  "capabilities": [
2348
+ "tools"
 
2349
  ],
2350
+ "hf_id": "moonshotai/Kimi-K2-Instruct",
2351
+ "size_b": 1026.5
2352
  },
2353
  {
2354
+ "name": "minimaxi/MiniMax-M2.7-highspeed",
2355
+ "type": "vision",
2356
  "input_price_per_1m": 0.6,
2357
+ "output_price_per_1m": 2.4,
2358
  "currency": "USD",
 
2359
  "capabilities": [
2360
+ "vision",
2361
  "reasoning"
2362
  ],
2363
+ "context_window": 200000,
2364
+ "hf_id": "MiniMaxAI/MiniMax-M2.7",
2365
+ "hf_private": true
2366
  },
2367
  {
2368
  "name": "minimaxi/MiniMax-M2.5-highspeed",
 
2379
  "size_b": 228.7
2380
  },
2381
  {
2382
+ "name": "moonshot/kimi-k2-0711-preview",
2383
+ "type": "chat",
2384
+ "input_price_per_1m": 0.6,
2385
+ "output_price_per_1m": 2.5,
2386
+ "currency": "USD",
2387
+ "context_window": 131072,
2388
+ "capabilities": [
2389
+ "tools"
2390
+ ],
2391
+ "hf_id": "moonshotai/Kimi-K2-Instruct",
2392
+ "size_b": 1026.5
2393
+ },
2394
+ {
2395
+ "name": "moonshot/kimi-k2.5",
2396
  "type": "vision",
2397
  "input_price_per_1m": 0.6,
2398
+ "output_price_per_1m": 3,
2399
  "currency": "USD",
2400
  "capabilities": [
2401
  "vision",
2402
  "reasoning"
2403
  ],
2404
+ "context_window": 262144,
2405
+ "hf_id": "moonshotai/Kimi-K2.5",
2406
+ "size_b": 1058.6
2407
  },
2408
  {
2409
  "name": "moonshot/kimi-k2-thinking-turbo",
 
2446
  "size_b": 1026.5
2447
  },
2448
  {
2449
+ "name": "nebius/zai-org/GLM-4.5",
2450
  "type": "chat",
2451
  "input_price_per_1m": 0.6,
2452
+ "output_price_per_1m": 2.2,
 
 
 
 
 
 
 
 
 
 
 
 
 
2453
  "currency": "USD",
2454
+ "context_window": 128000,
2455
  "capabilities": [
2456
+ "tools",
2457
  "reasoning"
2458
  ],
2459
+ "hf_id": "zai-org/GLM-4.5",
2460
+ "size_b": 358.3
 
2461
  },
2462
  {
2463
  "name": "zai/GLM-4.6",
 
2498
  "hf_id": "zai-org/GLM-4.5",
2499
  "size_b": 358.3
2500
  },
2501
+ {
2502
+ "name": "novita/zai-org/glm-4.6",
2503
+ "type": "chat",
2504
+ "input_price_per_1m": 0.6,
2505
+ "output_price_per_1m": 2.2,
2506
+ "currency": "USD",
2507
+ "context_window": 204800,
2508
+ "capabilities": [
2509
+ "tools",
2510
+ "reasoning"
2511
+ ],
2512
+ "hf_id": "zai-org/GLM-4.6",
2513
+ "size_b": 356.8
2514
+ },
2515
+ {
2516
+ "name": "novita/zai-org/glm-4.5",
2517
+ "type": "chat",
2518
+ "input_price_per_1m": 0.6,
2519
+ "output_price_per_1m": 2.2,
2520
+ "currency": "USD",
2521
+ "context_window": 131072,
2522
+ "capabilities": [
2523
+ "tools",
2524
+ "reasoning"
2525
+ ],
2526
+ "hf_id": "zai-org/GLM-4.5",
2527
+ "size_b": 358.3
2528
+ },
2529
+ {
2530
+ "name": "deepinfra/zai-org/GLM-4.5",
2531
+ "type": "chat",
2532
+ "input_price_per_1m": 0.6,
2533
+ "output_price_per_1m": 2.2,
2534
+ "currency": "USD",
2535
+ "context_window": 131072,
2536
+ "capabilities": [
2537
+ "tools",
2538
+ "reasoning"
2539
+ ],
2540
+ "hf_id": "zai-org/GLM-4.5",
2541
+ "size_b": 358.3
2542
+ },
2543
+ {
2544
+ "name": "alibaba/qwen3.5",
2545
+ "type": "vision",
2546
+ "input_price_per_1m": 0.6,
2547
+ "output_price_per_1m": 3.6,
2548
+ "currency": "USD",
2549
+ "capabilities": [
2550
+ "vision",
2551
+ "reasoning"
2552
+ ],
2553
+ "context_window": 256000,
2554
+ "size_b": 9,
2555
+ "hf_id": "Qwen/Qwen3.5-9B"
2556
+ },
2557
  {
2558
  "name": "novita/microsoft/wizardlm-2-8x22b",
2559
  "type": "chat",
 
2577
  "hf_private": true
2578
  },
2579
  {
2580
+ "name": "novita/deepseek/deepseek-r1-turbo",
2581
+ "type": "chat",
2582
  "input_price_per_1m": 0.7,
2583
+ "output_price_per_1m": 2.5,
2584
  "currency": "USD",
2585
+ "context_window": 64000,
2586
  "capabilities": [
2587
+ "tools",
2588
+ "reasoning"
2589
  ],
2590
+ "hf_id": "deepseek-ai/DeepSeek-R1",
2591
+ "size_b": 671
2592
  },
2593
  {
2594
  "name": "novita/deepseek/deepseek-prover-v2-671b",
 
2600
  "size_b": 671
2601
  },
2602
  {
2603
+ "name": "parasail/parasail-qwen25-vl-72b-instruct",
2604
+ "type": "vision",
2605
  "input_price_per_1m": 0.7,
2606
+ "output_price_per_1m": 0.7,
2607
  "currency": "USD",
 
2608
  "capabilities": [
2609
+ "vision"
 
2610
  ],
2611
+ "context_window": 32768,
2612
+ "size_b": 72
2613
  },
2614
  {
2615
+ "name": "openai/gpt-5.4-mini",
2616
  "type": "vision",
2617
  "input_price_per_1m": 0.75,
2618
  "output_price_per_1m": 4.5,
 
2625
  "hf_private": true
2626
  },
2627
  {
2628
+ "name": "openai-responses/gpt-5.4-mini",
2629
  "type": "vision",
2630
  "input_price_per_1m": 0.75,
2631
  "output_price_per_1m": 4.5,
 
2638
  "hf_private": true
2639
  },
2640
  {
2641
+ "name": "nebius/deepseek-ai/DeepSeek-R1-0528",
2642
  "type": "chat",
2643
  "input_price_per_1m": 0.8,
2644
+ "output_price_per_1m": 2.4,
2645
  "currency": "USD",
2646
+ "context_window": 164000,
2647
+ "capabilities": [
2648
+ "tools",
2649
+ "reasoning"
2650
+ ],
2651
+ "hf_id": "deepseek-ai/DeepSeek-R1-0528",
2652
+ "size_b": 684.5
2653
  },
2654
  {
2655
  "name": "novita/deepseek/deepseek-r1-distill-llama-70b",
 
2678
  "hf_id": "Qwen/Qwen2.5-VL-72B-Instruct"
2679
  },
2680
  {
2681
+ "name": "deepinfra/meta-llama/Meta-Llama-3.1-405B-Instruct",
2682
  "type": "chat",
2683
  "input_price_per_1m": 0.8,
2684
+ "output_price_per_1m": 0.8,
2685
  "currency": "USD",
2686
+ "context_window": 130815,
2687
+ "size_b": 405
 
 
 
 
 
2688
  },
2689
  {
2690
  "name": "deepinfra/deepseek-ai/DeepSeek-R1",
 
2698
  "reasoning"
2699
  ],
2700
  "hf_id": "deepseek-ai/DeepSeek-R1",
2701
+ "size_b": 671
2702
  },
2703
  {
2704
  "name": "deepinfra/deepseek-ai/DeepSeek-V3",
 
2760
  "reasoning"
2761
  ],
2762
  "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
2763
+ "size_b": 671
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2764
  },
2765
  {
2766
+ "name": "parasail/parasail-kimi-k2-instruct",
2767
+ "type": "chat",
2768
+ "input_price_per_1m": 0.99,
2769
+ "output_price_per_1m": 2.99,
2770
  "currency": "USD",
2771
  "context_window": 131072,
 
 
 
2772
  "hf_private": true
2773
  },
2774
  {
 
2797
  "hf_private": true
2798
  },
2799
  {
2800
+ "name": "anthropic/claude-haiku-4-5",
 
 
 
 
 
 
 
 
 
 
 
 
 
2801
  "type": "vision",
2802
  "input_price_per_1m": 1,
2803
  "output_price_per_1m": 5,
 
2849
  "size_b": 1026.5
2850
  },
2851
  {
2852
+ "name": "novita/GLM-5",
2853
+ "type": "chat",
2854
+ "input_price_per_1m": 1,
2855
+ "output_price_per_1m": 3.2,
2856
  "currency": "USD",
2857
  "capabilities": [
2858
  "reasoning"
2859
  ],
2860
+ "context_window": 202800,
2861
+ "hf_id": "zai-org/GLM-5",
2862
+ "size_b": 753.9
2863
+ },
2864
+ {
2865
+ "name": "perplexity/sonar",
2866
+ "type": "vision",
2867
+ "input_price_per_1m": 1,
2868
+ "output_price_per_1m": 1,
2869
+ "currency": "USD",
2870
+ "context_window": 131072,
2871
+ "capabilities": [
2872
+ "vision"
2873
+ ],
2874
+ "hf_private": true
2875
+ },
2876
+ {
2877
+ "name": "bedrock/claude-haiku-4-5",
2878
+ "type": "vision",
2879
+ "input_price_per_1m": 1,
2880
+ "output_price_per_1m": 5,
2881
+ "currency": "USD",
2882
+ "capabilities": [
2883
+ "vision"
2884
+ ],
2885
  "context_window": 200000,
2886
  "hf_private": true
2887
  },
2888
  {
2889
+ "name": "alibaba/qwen3-coder-plus",
2890
+ "type": "vision",
2891
+ "input_price_per_1m": 1,
2892
+ "output_price_per_1m": 5,
2893
+ "currency": "USD",
2894
+ "capabilities": [
2895
+ "vision"
2896
+ ],
2897
+ "context_window": 1048576,
2898
+ "hf_private": true
2899
+ },
2900
+ {
2901
+ "name": "openai/o3-mini",
2902
  "type": "chat",
2903
  "input_price_per_1m": 1.1,
2904
  "output_price_per_1m": 4.4,
 
2922
  "hf_private": true
2923
  },
2924
  {
2925
+ "name": "openai-responses/o3-mini",
2926
  "type": "chat",
2927
  "input_price_per_1m": 1.1,
2928
  "output_price_per_1m": 4.4,
 
2934
  "hf_private": true
2935
  },
2936
  {
2937
+ "name": "azure/o4-mini",
2938
+ "type": "vision",
2939
+ "input_price_per_1m": 1.1,
2940
+ "output_price_per_1m": 4.4,
2941
  "currency": "USD",
2942
+ "capabilities": [
2943
+ "reasoning"
2944
+ ],
2945
+ "context_window": 200000,
2946
+ "hf_private": true
2947
  },
2948
  {
2949
  "name": "moonshot/kimi-k2-turbo-preview",
 
2959
  "size_b": 1026.5
2960
  },
2961
  {
2962
+ "name": "together/Qwen/Qwen2.5-72B-Instruct-Turbo",
2963
+ "type": "chat",
2964
+ "input_price_per_1m": 1.2,
2965
+ "output_price_per_1m": 1.2,
2966
+ "currency": "USD",
2967
+ "context_window": 32768,
2968
+ "size_b": 72
2969
+ },
2970
+ {
2971
+ "name": "vertex/gemini-2.5-pro",
2972
  "type": "vision",
2973
  "input_price_per_1m": 1.25,
2974
  "output_price_per_1m": 10,
2975
  "currency": "USD",
2976
  "capabilities": [
2977
+ "vision",
2978
  "reasoning"
2979
  ],
2980
+ "context_window": 1048576,
2981
  "hf_private": true
2982
  },
2983
  {
2984
+ "name": "openai/gpt-5-chat",
2985
  "type": "vision",
2986
  "input_price_per_1m": 1.25,
2987
  "output_price_per_1m": 10,
2988
  "currency": "USD",
2989
  "capabilities": [
2990
+ "vision",
2991
  "reasoning"
2992
  ],
2993
+ "context_window": 128000,
2994
  "hf_private": true
2995
  },
2996
  {
2997
+ "name": "openai/gpt-5.1-chat",
2998
+ "type": "vision",
2999
  "input_price_per_1m": 1.25,
3000
+ "output_price_per_1m": 10,
3001
  "currency": "USD",
 
3002
  "capabilities": [
3003
+ "vision",
3004
  "reasoning"
3005
  ],
3006
+ "context_window": 128000,
3007
+ "hf_private": true
3008
  },
3009
  {
3010
+ "name": "openai/gpt-5.1",
3011
  "type": "vision",
3012
  "input_price_per_1m": 1.25,
3013
  "output_price_per_1m": 10,
 
3020
  "hf_private": true
3021
  },
3022
  {
3023
+ "name": "google/gemini-2.5-pro",
3024
  "type": "vision",
3025
  "input_price_per_1m": 1.25,
3026
  "output_price_per_1m": 10,
 
3029
  "vision",
3030
  "reasoning"
3031
  ],
3032
+ "context_window": 1048576,
3033
  "hf_private": true
3034
  },
3035
  {
3036
+ "name": "together/deepseek-ai/DeepSeek-V3",
3037
+ "type": "chat",
3038
  "input_price_per_1m": 1.25,
3039
+ "output_price_per_1m": 1.25,
3040
  "currency": "USD",
3041
+ "context_window": 131072,
3042
  "capabilities": [
3043
+ "tools",
3044
  "reasoning"
3045
  ],
3046
+ "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
3047
+ "size_b": 671
3048
  },
3049
  {
3050
+ "name": "openai-responses/gpt-5.1-codex",
3051
  "type": "vision",
3052
  "input_price_per_1m": 1.25,
3053
  "output_price_per_1m": 10,
 
3060
  "hf_private": true
3061
  },
3062
  {
3063
+ "name": "openai-responses/gpt-5-codex",
3064
  "type": "vision",
3065
  "input_price_per_1m": 1.25,
3066
  "output_price_per_1m": 10,
 
3069
  "vision",
3070
  "reasoning"
3071
  ],
3072
+ "context_window": 400000,
3073
  "hf_private": true
3074
  },
3075
  {
3076
+ "name": "openai-responses/gpt-5",
3077
  "type": "vision",
3078
  "input_price_per_1m": 1.25,
3079
  "output_price_per_1m": 10,
 
3086
  "hf_private": true
3087
  },
3088
  {
3089
+ "name": "openai-responses/gpt-5.1",
3090
  "type": "vision",
3091
  "input_price_per_1m": 1.25,
3092
  "output_price_per_1m": 10,
 
3095
  "vision",
3096
  "reasoning"
3097
  ],
3098
+ "context_window": 400000,
3099
  "hf_private": true
3100
  },
3101
  {
3102
+ "name": "azure/gpt-5",
3103
  "type": "vision",
3104
  "input_price_per_1m": 1.25,
3105
  "output_price_per_1m": 10,
3106
  "currency": "USD",
3107
  "capabilities": [
 
3108
  "reasoning"
3109
  ],
3110
+ "context_window": 200000,
3111
  "hf_private": true
3112
  },
3113
  {
3114
+ "name": "azure/gpt-5.1",
3115
  "type": "vision",
3116
  "input_price_per_1m": 1.25,
3117
  "output_price_per_1m": 10,
3118
  "currency": "USD",
3119
  "capabilities": [
 
3120
  "reasoning"
3121
  ],
3122
+ "context_window": 200000,
3123
  "hf_private": true
3124
  },
3125
  {
3126
+ "name": "coding/gemini-2.5-pro",
3127
  "type": "vision",
3128
  "input_price_per_1m": 1.25,
3129
  "output_price_per_1m": 10,
 
3136
  "hf_private": true
3137
  },
3138
  {
3139
+ "name": "novita/sao10k/l3-70b-euryale-v2.1",
3140
  "type": "chat",
3141
  "input_price_per_1m": 1.48,
3142
  "output_price_per_1m": 1.48,
 
3145
  "size_b": 70
3146
  },
3147
  {
3148
+ "name": "novita/sao10k/l31-70b-euryale-v2.2",
3149
  "type": "chat",
3150
  "input_price_per_1m": 1.48,
3151
  "output_price_per_1m": 1.48,
 
3166
  "hf_private": true
3167
  },
3168
  {
3169
+ "name": "openai/gpt-5.3-chat",
3170
  "type": "vision",
3171
  "input_price_per_1m": 1.75,
3172
  "output_price_per_1m": 14,
 
3175
  "vision",
3176
  "reasoning"
3177
  ],
3178
+ "context_window": 128000,
3179
  "hf_private": true
3180
  },
3181
  {
3182
+ "name": "openai/gpt-5.2-chat",
3183
  "type": "vision",
3184
  "input_price_per_1m": 1.75,
3185
  "output_price_per_1m": 14,
 
3188
  "vision",
3189
  "reasoning"
3190
  ],
3191
+ "context_window": 128000,
3192
  "hf_private": true
3193
  },
3194
  {
3195
+ "name": "openai/gpt-5.2",
3196
  "type": "vision",
3197
  "input_price_per_1m": 1.75,
3198
  "output_price_per_1m": 14,
3199
  "currency": "USD",
3200
  "capabilities": [
3201
+ "vision",
3202
  "reasoning"
3203
  ],
3204
+ "context_window": 400000,
3205
  "hf_private": true
3206
  },
3207
  {
 
3218
  "hf_private": true
3219
  },
3220
  {
3221
+ "name": "openai-responses/gpt-5.3-codex",
3222
  "type": "vision",
3223
  "input_price_per_1m": 1.75,
3224
  "output_price_per_1m": 14,
 
3231
  "hf_private": true
3232
  },
3233
  {
3234
+ "name": "openai-responses/gpt-5.2-codex",
3235
  "type": "vision",
3236
  "input_price_per_1m": 1.75,
3237
  "output_price_per_1m": 14,
 
3244
  "hf_private": true
3245
  },
3246
  {
3247
+ "name": "openai-responses/gpt-5.2",
3248
  "type": "vision",
3249
  "input_price_per_1m": 1.75,
3250
  "output_price_per_1m": 14,
 
3257
  "hf_private": true
3258
  },
3259
  {
3260
+ "name": "azure/gpt-5.3-codex",
3261
  "type": "vision",
3262
  "input_price_per_1m": 1.75,
3263
  "output_price_per_1m": 14,
 
3270
  "hf_private": true
3271
  },
3272
  {
3273
+ "name": "azure/openai-responses/gpt-5.2-codex",
3274
  "type": "vision",
3275
  "input_price_per_1m": 1.75,
3276
  "output_price_per_1m": 14,
 
3279
  "vision",
3280
  "reasoning"
3281
  ],
3282
+ "context_window": 400000,
3283
  "hf_private": true
3284
  },
3285
  {
3286
+ "name": "azure/gpt-5.2",
3287
  "type": "vision",
3288
  "input_price_per_1m": 1.75,
3289
  "output_price_per_1m": 14,
3290
  "currency": "USD",
3291
  "capabilities": [
 
3292
  "reasoning"
3293
  ],
3294
+ "context_window": 200000,
3295
  "hf_private": true
3296
  },
3297
  {
3298
+ "name": "mistral/pixtral-large-latest",
3299
+ "type": "chat",
3300
+ "input_price_per_1m": 2,
3301
+ "output_price_per_1m": 5,
3302
+ "currency": "USD",
3303
+ "context_window": 131072,
3304
+ "hf_private": true
3305
+ },
3306
+ {
3307
+ "name": "vertex/gemini-3-pro-preview",
3308
  "type": "vision",
3309
  "input_price_per_1m": 2,
3310
+ "output_price_per_1m": 12,
3311
  "currency": "USD",
3312
  "capabilities": [
3313
+ "vision",
3314
+ "reasoning"
3315
  ],
3316
+ "context_window": 1048576,
3317
  "hf_private": true
3318
  },
3319
  {
3320
+ "name": "vertex/gemini-3.1-pro-preview",
3321
  "type": "vision",
3322
  "input_price_per_1m": 2,
3323
+ "output_price_per_1m": 12,
3324
  "currency": "USD",
3325
  "capabilities": [
3326
+ "vision",
3327
+ "reasoning"
3328
  ],
3329
+ "context_window": 1048576,
3330
  "hf_private": true
3331
  },
3332
  {
3333
+ "name": "vertex/gemini-3-pro-image-preview",
3334
  "type": "vision",
3335
  "input_price_per_1m": 2,
3336
+ "output_price_per_1m": 12,
3337
  "currency": "USD",
 
3338
  "capabilities": [
3339
  "vision",
3340
  "reasoning"
3341
  ],
3342
+ "context_window": 1048576,
3343
  "hf_private": true
3344
  },
3345
  {
3346
+ "name": "openai/gpt-4.1",
3347
  "type": "vision",
3348
  "input_price_per_1m": 2,
3349
  "output_price_per_1m": 8,
 
3355
  "hf_private": true
3356
  },
3357
  {
3358
+ "name": "openai/o4-mini-deep-research",
3359
+ "type": "vision",
3360
+ "input_price_per_1m": 2,
3361
+ "output_price_per_1m": 8,
3362
+ "currency": "USD",
3363
+ "capabilities": [
3364
+ "vision",
3365
+ "reasoning"
3366
+ ],
3367
+ "context_window": 100000,
3368
+ "hf_private": true
3369
+ },
3370
+ {
3371
+ "name": "google/gemini-3-pro-preview",
3372
  "type": "vision",
3373
  "input_price_per_1m": 2,
3374
  "output_price_per_1m": 12,
 
3381
  "hf_private": true
3382
  },
3383
  {
3384
+ "name": "google/gemini-3.1-pro-preview",
3385
  "type": "vision",
3386
  "input_price_per_1m": 2,
3387
  "output_price_per_1m": 12,
 
3394
  "hf_private": true
3395
  },
3396
  {
3397
+ "name": "nebius/deepseek-ai/DeepSeek-V3-0324-fast",
3398
+ "type": "chat",
3399
  "input_price_per_1m": 2,
3400
+ "output_price_per_1m": 6,
3401
  "currency": "USD",
3402
+ "context_window": 128000,
 
 
 
 
3403
  "hf_private": true
3404
  },
3405
  {
3406
+ "name": "openai-responses/gpt-4.1",
3407
  "type": "vision",
3408
  "input_price_per_1m": 2,
3409
  "output_price_per_1m": 8,
 
3415
  "hf_private": true
3416
  },
3417
  {
3418
+ "name": "azure/gpt-4.1",
3419
  "type": "vision",
3420
  "input_price_per_1m": 2,
3421
  "output_price_per_1m": 8,
3422
  "currency": "USD",
3423
  "capabilities": [
3424
+ "vision"
 
3425
  ],
3426
+ "context_window": 1047576,
3427
  "hf_private": true
3428
  },
3429
  {
3430
+ "name": "azure/openai-responses/gpt-4.1",
3431
  "type": "vision",
3432
  "input_price_per_1m": 2,
3433
+ "output_price_per_1m": 8,
3434
  "currency": "USD",
3435
  "capabilities": [
3436
+ "vision"
 
3437
  ],
3438
+ "context_window": 1047576,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3439
  "hf_private": true
3440
  },
3441
  {
3442
+ "name": "perplexity/sonar-reasoning-pro",
3443
  "type": "vision",
3444
  "input_price_per_1m": 2,
3445
+ "output_price_per_1m": 8,
3446
  "currency": "USD",
3447
+ "context_window": 131072,
3448
  "capabilities": [
3449
  "vision",
3450
  "reasoning"
3451
  ],
 
3452
  "hf_private": true
3453
  },
3454
  {
3455
+ "name": "xai/grok-4.2-beta",
3456
  "type": "vision",
3457
  "input_price_per_1m": 2,
3458
+ "output_price_per_1m": 6,
3459
  "currency": "USD",
3460
  "capabilities": [
3461
  "vision",
3462
  "reasoning"
3463
  ],
3464
+ "context_window": 2000000,
3465
  "hf_private": true
3466
  },
3467
  {
3468
+ "name": "xai/grok-2-1212",
3469
  "type": "chat",
3470
  "input_price_per_1m": 2,
3471
+ "output_price_per_1m": 10,
3472
  "currency": "USD",
3473
  "context_window": 131072,
3474
  "hf_private": true
3475
  },
3476
  {
3477
+ "name": "openai/gpt-4o-2024-11-20",
3478
  "type": "vision",
3479
  "input_price_per_1m": 2.5,
3480
+ "output_price_per_1m": 10,
3481
  "currency": "USD",
3482
  "capabilities": [
3483
+ "vision"
 
3484
  ],
3485
+ "context_window": 128000,
3486
  "hf_private": true
3487
  },
3488
  {
3489
+ "name": "openai/gpt-4o",
3490
  "type": "vision",
3491
  "input_price_per_1m": 2.5,
3492
+ "output_price_per_1m": 10,
3493
  "currency": "USD",
3494
  "capabilities": [
3495
+ "vision"
 
3496
  ],
3497
+ "context_window": 128000,
3498
  "hf_private": true
3499
  },
3500
  {
 
3535
  "hf_private": true
3536
  },
3537
  {
3538
+ "name": "openai-responses/gpt-5.4",
3539
  "type": "vision",
3540
  "input_price_per_1m": 2.5,
3541
+ "output_price_per_1m": 15,
3542
  "currency": "USD",
3543
  "capabilities": [
3544
+ "vision",
3545
+ "reasoning"
3546
  ],
3547
+ "context_window": 1050000,
3548
  "hf_private": true
3549
  },
3550
  {
3551
+ "name": "azure/openai-responses/gpt-5.4",
3552
  "type": "vision",
3553
  "input_price_per_1m": 2.5,
3554
+ "output_price_per_1m": 15,
3555
  "currency": "USD",
3556
  "capabilities": [
3557
+ "vision",
3558
+ "reasoning"
3559
  ],
3560
+ "context_window": 1050000,
3561
  "hf_private": true
3562
  },
3563
  {
3564
+ "name": "vertex/claude-sonnet-4",
3565
  "type": "vision",
3566
  "input_price_per_1m": 3,
3567
  "output_price_per_1m": 15,
 
3570
  "vision",
3571
  "reasoning"
3572
  ],
3573
+ "context_window": 200000,
3574
  "hf_private": true
3575
  },
3576
  {
3577
+ "name": "vertex/claude-sonnet-4-5",
3578
  "type": "vision",
3579
  "input_price_per_1m": 3,
3580
  "output_price_per_1m": 15,
 
3583
  "vision",
3584
  "reasoning"
3585
  ],
3586
+ "context_window": 200000,
3587
  "hf_private": true
3588
  },
3589
  {
3590
+ "name": "vertex/claude-3-7-sonnet",
3591
  "type": "vision",
3592
  "input_price_per_1m": 3,
3593
  "output_price_per_1m": 15,
 
3596
  "vision",
3597
  "reasoning"
3598
  ],
3599
+ "context_window": 200000,
3600
  "hf_private": true
3601
  },
3602
  {
3603
+ "name": "vertex/claude-3-5-sonnet",
3604
+ "type": "vision",
3605
  "input_price_per_1m": 3,
3606
+ "output_price_per_1m": 15,
3607
  "currency": "USD",
 
3608
  "capabilities": [
3609
+ "vision"
 
3610
  ],
3611
+ "context_window": 200000,
3612
+ "hf_private": true
3613
  },
3614
  {
3615
+ "name": "anthropic/claude-sonnet-4-5",
3616
  "type": "vision",
3617
  "input_price_per_1m": 3,
3618
  "output_price_per_1m": 15,
3619
  "currency": "USD",
 
3620
  "capabilities": [
3621
+ "vision",
3622
+ "reasoning"
3623
  ],
3624
+ "context_window": 1000000,
3625
  "hf_private": true
3626
  },
3627
  {
3628
+ "name": "anthropic/claude-sonnet-4-6",
3629
  "type": "vision",
3630
  "input_price_per_1m": 3,
3631
  "output_price_per_1m": 15,
3632
  "currency": "USD",
3633
  "capabilities": [
3634
+ "vision",
3635
+ "reasoning"
3636
  ],
3637
+ "context_window": 1000000,
3638
  "hf_private": true
3639
  },
3640
  {
3641
+ "name": "anthropic/claude-sonnet-4",
3642
  "type": "vision",
3643
  "input_price_per_1m": 3,
3644
  "output_price_per_1m": 15,
 
3647
  "vision",
3648
  "reasoning"
3649
  ],
3650
+ "context_window": 1000000,
3651
  "hf_private": true
3652
  },
3653
  {
3654
+ "name": "together/deepseek-ai/DeepSeek-R1",
3655
+ "type": "chat",
3656
  "input_price_per_1m": 3,
3657
+ "output_price_per_1m": 7,
3658
  "currency": "USD",
3659
+ "context_window": 64000,
3660
  "capabilities": [
3661
+ "tools",
3662
  "reasoning"
3663
  ],
3664
+ "hf_id": "deepseek-ai/DeepSeek-R1",
3665
+ "size_b": 671
3666
  },
3667
  {
3668
+ "name": "coding/claude-sonnet-4-20250514",
3669
  "type": "vision",
3670
  "input_price_per_1m": 3,
3671
  "output_price_per_1m": 15,
 
3678
  "hf_private": true
3679
  },
3680
  {
3681
+ "name": "perplexity/sonar-pro",
3682
  "type": "vision",
3683
  "input_price_per_1m": 3,
3684
  "output_price_per_1m": 15,
3685
  "currency": "USD",
3686
+ "context_window": 204800,
3687
  "capabilities": [
3688
  "vision"
3689
  ],
 
3690
  "hf_private": true
3691
  },
3692
  {
3693
+ "name": "bedrock/claude-sonnet-4-5",
3694
  "type": "vision",
3695
  "input_price_per_1m": 3,
3696
  "output_price_per_1m": 15,
 
3716
  "hf_private": true
3717
  },
3718
  {
3719
+ "name": "bedrock/claude-sonnet-4-6",
3720
  "type": "vision",
3721
  "input_price_per_1m": 3,
3722
  "output_price_per_1m": 15,
 
3725
  "vision",
3726
  "reasoning"
3727
  ],
3728
+ "context_window": 1000000,
3729
  "hf_private": true
3730
  },
3731
  {
3732
+ "name": "bedrock/claude-sonnet-4",
3733
  "type": "vision",
3734
  "input_price_per_1m": 3,
3735
  "output_price_per_1m": 15,
 
3738
  "vision",
3739
  "reasoning"
3740
  ],
3741
+ "context_window": 200000,
3742
  "hf_private": true
3743
  },
3744
  {
3745
+ "name": "xai/grok-4",
3746
  "type": "vision",
3747
  "input_price_per_1m": 3,
3748
  "output_price_per_1m": 15,
3749
  "currency": "USD",
3750
  "capabilities": [
3751
+ "vision"
 
3752
  ],
3753
+ "context_window": 256000,
3754
  "hf_private": true
3755
  },
3756
  {
 
3765
  "reasoning"
3766
  ],
3767
  "hf_id": "deepseek-ai/DeepSeek-R1",
3768
+ "size_b": 671
3769
  },
3770
  {
3771
+ "name": "vertex/claude-opus-4-5",
3772
  "type": "vision",
3773
  "input_price_per_1m": 5,
3774
  "output_price_per_1m": 25,
 
3781
  "hf_private": true
3782
  },
3783
  {
3784
+ "name": "vertex/claude-opus-4-6",
3785
  "type": "vision",
3786
  "input_price_per_1m": 5,
3787
  "output_price_per_1m": 25,
 
3794
  "hf_private": true
3795
  },
3796
  {
3797
+ "name": "openai/chatgpt-4o",
3798
  "type": "vision",
3799
  "input_price_per_1m": 5,
3800
+ "output_price_per_1m": 15,
3801
  "currency": "USD",
3802
  "capabilities": [
3803
+ "vision"
 
3804
  ],
3805
+ "context_window": 128000,
3806
  "hf_private": true
3807
  },
3808
  {
3809
+ "name": "anthropic/claude-opus-4-6",
3810
  "type": "vision",
3811
  "input_price_per_1m": 5,
3812
  "output_price_per_1m": 25,
 
3815
  "vision",
3816
  "reasoning"
3817
  ],
3818
+ "context_window": 1000000,
3819
  "hf_private": true
3820
  },
3821
  {
3822
+ "name": "anthropic/claude-opus-4-5",
3823
  "type": "vision",
3824
  "input_price_per_1m": 5,
 
 
 
 
 
 
 
 
 
 
 
 
3825
  "output_price_per_1m": 25,
3826
  "currency": "USD",
 
3827
  "capabilities": [
3828
+ "vision",
3829
+ "reasoning"
3830
  ],
3831
+ "context_window": 200000,
3832
  "hf_private": true
3833
  },
3834
  {
 
3857
  "context_window": 1000000,
3858
  "hf_private": true
3859
  },
3860
+ {
3861
+ "name": "xai/grok-3",
3862
+ "type": "chat",
3863
+ "input_price_per_1m": 5,
3864
+ "output_price_per_1m": 25,
3865
+ "currency": "USD",
3866
+ "context_window": 131072,
3867
+ "capabilities": [
3868
+ "tools"
3869
+ ],
3870
+ "hf_private": true
3871
+ },
3872
  {
3873
  "name": "openai/o3-deep-research",
3874
  "type": "vision",
 
3883
  "hf_private": true
3884
  },
3885
  {
3886
+ "name": "vertex/claude-opus-4",
3887
  "type": "vision",
3888
  "input_price_per_1m": 15,
3889
  "output_price_per_1m": 75,
 
3896
  "hf_private": true
3897
  },
3898
  {
3899
+ "name": "vertex/claude-opus-4-1",
3900
  "type": "vision",
3901
  "input_price_per_1m": 15,
3902
  "output_price_per_1m": 75,
 
3909
  "hf_private": true
3910
  },
3911
  {
3912
+ "name": "openai/o1",
3913
  "type": "vision",
3914
  "input_price_per_1m": 15,
3915
+ "output_price_per_1m": 60,
3916
  "currency": "USD",
3917
  "capabilities": [
3918
  "vision",
3919
  "reasoning"
3920
  ],
3921
+ "context_window": 200000,
3922
  "hf_private": true
3923
  },
3924
  {
3925
+ "name": "anthropic/claude-opus-4",
3926
  "type": "vision",
3927
  "input_price_per_1m": 15,
3928
  "output_price_per_1m": 75,
 
3935
  "hf_private": true
3936
  },
3937
  {
3938
+ "name": "anthropic/claude-opus-4-1",
3939
  "type": "vision",
3940
  "input_price_per_1m": 15,
3941
  "output_price_per_1m": 75,
 
3948
  "hf_private": true
3949
  },
3950
  {
3951
+ "name": "openai-responses/gpt-5-pro",
3952
  "type": "vision",
3953
  "input_price_per_1m": 15,
3954
+ "output_price_per_1m": 120,
3955
  "currency": "USD",
3956
  "capabilities": [
3957
  "vision",
3958
  "reasoning"
3959
  ],
3960
+ "context_window": 400000,
3961
  "hf_private": true
3962
  },
3963
  {
 
4000
  "hf_private": true
4001
  },
4002
  {
4003
+ "name": "openai/gpt-5.4-pro",
4004
  "type": "vision",
4005
  "input_price_per_1m": 30,
4006
  "output_price_per_1m": 180,
 
4026
  "hf_private": true
4027
  },
4028
  {
4029
+ "name": "azure/openai-responses/gpt-5.4-pro",
4030
  "type": "vision",
4031
  "input_price_per_1m": 30,
4032
  "output_price_per_1m": 180,
 
4348
  "input_price_per_1m": 0.75,
4349
  "output_price_per_1m": 2.25,
4350
  "currency": "USD",
4351
+ "flavor": "fast",
4352
+ "size_b": 671,
4353
+ "hf_id": "deepseek-ai/DeepSeek-V3"
4354
  },
4355
  {
4356
  "name": "DeepSeek-V3-0324 (base)",
 
5565
  "input_price_per_1m": 0.065,
5566
  "output_price_per_1m": 0.14,
5567
  "currency": "USD",
5568
+ "hf_id": "microsoft/phi-4",
5569
+ "size_b": 14
5570
  },
5571
  {
5572
  "name": "baidu/ernie-4.5-21b-a3b-thinking",
 
7152
  "hf_id": "deepseek-ai/DeepSeek-V3",
7153
  "capabilities": [
7154
  "tools"
7155
+ ],
7156
+ "size_b": 671
7157
  },
7158
  {
7159
  "name": "qwen/qwen3.5-397b-a17b",
scripts/fetch-providers.js CHANGED
@@ -53,7 +53,7 @@ function updateProviderModels(providers, providerName, models) {
53
  return false;
54
  }
55
 
56
- // Smart merge: preserve existing metadata (size_b, hf_id, ollama_id, capabilities, hf_private) if missing in new data
57
  const existingMap = new Map((provider.models || []).map(m => [m.name, m]));
58
 
59
  provider.models = models.map(newModel => {
@@ -165,9 +165,6 @@ function estimateParams(config) {
165
 
166
  if (h && l && v) {
167
  // Basic transformer param estimation: Layers * (Embedding + Attention + MLP)
168
- // Embedding: v * h
169
- // Attention: 4 * h^2
170
- // MLP: 2 * h * i (or 8 * h^2 roughly if i missing)
171
  const intermediate = i || (4 * h);
172
  const params = (v * h) + l * (4 * (h * h) + 2 * (h * intermediate));
173
  return params;
@@ -250,17 +247,70 @@ const MANUAL_HF_ID_MAP = {
250
  'bge m3': 'BAAI/bge-m3',
251
  'lightonocr 2': 'lightonai/LightOnOCR-2-1B',
252
  'flux 1 schnell': 'black-forest-labs/FLUX.1-schnell',
 
253
  'paraphrase multilingual mpnet base v2': 'sentence-transformers/paraphrase-multilingual-mpnet-base-v2',
254
  'bge large en v1 5': 'BAAI/bge-large-en-v1.5',
255
  'bge multilingual gemma2': 'BAAI/bge-multilingual-gemma2',
256
  'photomaker v2': 'TencentARC/PhotoMaker-V2',
257
- 'flux schnell': 'black-forest-labs/FLUX.1-schnell',
258
  // Qwen mappings
 
 
 
 
 
 
 
 
 
 
259
  'qwen3 coder flash': 'Qwen/Qwen2.5-Coder-7B-Instruct',
 
260
  'qwen3 coder plus': 'Qwen/Qwen2.5-Coder-32B-Instruct',
261
  'qwen 3 5 flash': 'Qwen/Qwen2.5-7B-Instruct',
 
262
  'qwen vl plus': 'Qwen/Qwen2-VL-7B-Instruct',
263
  'qwen vl max': 'Qwen/Qwen2-VL-72B-Instruct',
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
264
  // FLUX detailed mappings
265
  'flux 1 dev': 'black-forest-labs/FLUX.1-dev',
266
  'flux dev': 'black-forest-labs/FLUX.1-dev',
@@ -292,7 +342,21 @@ const MANUAL_SIZE_MAP = {
292
  'black-forest-labs/FLUX.2-klein-9B': 9,
293
  // Mistral family
294
  'mistralai/Mistral-Large-Instruct-2407': 123,
295
- 'mistralai/Mistral-Large-Instruct-2411': 675, // 41B active
 
 
 
 
 
 
 
 
 
 
 
 
 
 
296
  };
297
 
298
  // Propagate capabilities and size from benchmarks, OpenRouter, or HF Hub to all other providers' models.
@@ -334,21 +398,24 @@ async function propagateExtraData(data) {
334
  const n = normName(model.name);
335
 
336
  // 0. MANUAL OVERRIDE: Link common models to their HF IDs
337
- if (!model.hf_id && MANUAL_HF_ID_MAP[n]) {
338
- model.hf_id = MANUAL_HF_ID_MAP[n];
 
 
 
 
339
  }
340
 
341
- // 1. STRUCTURED LOOKUP: Match size by hf_id if available (Benchmark gold-standard)
342
- if (model.hf_id) {
343
- // High-confidence manual override (always overwrite even if size already exists)
344
- if (MANUAL_SIZE_MAP[model.hf_id]) {
345
  model.size_b = MANUAL_SIZE_MAP[model.hf_id];
346
  propagatedSize++;
347
- } else if (!model.size_b) {
348
- // Fallback to benchmarks if size missing
349
- const size = hfIdToSize.get(model.hf_id.toLowerCase());
350
- if (size) { model.size_b = size; propagatedSize++; }
351
  }
 
 
 
 
352
  }
353
 
354
 
@@ -401,12 +468,11 @@ async function propagateExtraData(data) {
401
  }
402
 
403
  // 6. QUEUE: Still missing size? Try Hub API or Ollama
404
- if (!model.size_b) {
405
- if (!model.hf_private && (model.name.includes('/') || model.hf_id)) {
406
- hfLookupQueue.push(model);
407
- } else if (model.type === 'chat') {
408
- ollamaLookupQueue.push(model);
409
- }
410
  }
411
  }
412
  }
 
53
  return false;
54
  }
55
 
56
+ // Smart merge: preserve existing metadata (size_b, hf_id, ollama_id, capabilities, hf_private)
57
  const existingMap = new Map((provider.models || []).map(m => [m.name, m]));
58
 
59
  provider.models = models.map(newModel => {
 
165
 
166
  if (h && l && v) {
167
  // Basic transformer param estimation: Layers * (Embedding + Attention + MLP)
 
 
 
168
  const intermediate = i || (4 * h);
169
  const params = (v * h) + l * (4 * (h * h) + 2 * (h * intermediate));
170
  return params;
 
247
  'bge m3': 'BAAI/bge-m3',
248
  'lightonocr 2': 'lightonai/LightOnOCR-2-1B',
249
  'flux 1 schnell': 'black-forest-labs/FLUX.1-schnell',
250
+ 'flux schnell': 'black-forest-labs/FLUX.1-schnell',
251
  'paraphrase multilingual mpnet base v2': 'sentence-transformers/paraphrase-multilingual-mpnet-base-v2',
252
  'bge large en v1 5': 'BAAI/bge-large-en-v1.5',
253
  'bge multilingual gemma2': 'BAAI/bge-multilingual-gemma2',
254
  'photomaker v2': 'TencentARC/PhotoMaker-V2',
 
255
  // Qwen mappings
256
+ 'qwen turbo': 'Alibaba/Qwen-Turbo',
257
+ 'alibaba qwen turbo': 'Alibaba/Qwen-Turbo',
258
+ 'qwen qwen turbo': 'Alibaba/Qwen-Turbo',
259
+ 'qwen plus': 'Alibaba/Qwen-Plus',
260
+ 'alibaba qwen plus': 'Alibaba/Qwen-Plus',
261
+ 'qwen qwen plus': 'Alibaba/Qwen-Plus',
262
+ 'qwen max': 'Alibaba/Qwen-Max',
263
+ 'alibaba qwen max': 'Alibaba/Qwen-Max',
264
+ 'qwen qwen max': 'Alibaba/Qwen-Max',
265
+ 'qwen 3 coder flash': 'Qwen/Qwen2.5-Coder-7B-Instruct',
266
  'qwen3 coder flash': 'Qwen/Qwen2.5-Coder-7B-Instruct',
267
+ 'qwen 3 coder plus': 'Qwen/Qwen2.5-Coder-32B-Instruct',
268
  'qwen3 coder plus': 'Qwen/Qwen2.5-Coder-32B-Instruct',
269
  'qwen 3 5 flash': 'Qwen/Qwen2.5-7B-Instruct',
270
+ 'qwen3 5 flash 02 23': 'Qwen/Qwen2.5-7B-Instruct',
271
  'qwen vl plus': 'Qwen/Qwen2-VL-7B-Instruct',
272
  'qwen vl max': 'Qwen/Qwen2-VL-72B-Instruct',
273
+ // DeepSeek mappings
274
+ 'deepseek chat': 'deepseek-ai/DeepSeek-V3',
275
+ 'deepseek reasoner': 'deepseek-ai/DeepSeek-R1',
276
+ 'deepseek v3 turbo': 'deepseek-ai/DeepSeek-V3',
277
+ 'deepseek v3 0324 fast': 'deepseek-ai/DeepSeek-V3',
278
+ 'deepseek r1t2 chimera': 'deepseek-ai/DeepSeek-R1',
279
+ 'deepseek v3 2 exp': 'deepseek-ai/DeepSeek-V3',
280
+ 'deepseek v3 2 speciale': 'deepseek-ai/DeepSeek-V3',
281
+ 'novita deepseek deepseek v3 turbo': 'deepseek-ai/DeepSeek-V3',
282
+ 'novita deepseek deepseek v3 0324': 'deepseek-ai/DeepSeek-V3',
283
+ 'novita deepseek deepseek chat': 'deepseek-ai/DeepSeek-V3',
284
+ 'deepseek deepseek chat': 'deepseek-ai/DeepSeek-V3',
285
+ 'deepseek deepseek reasoner': 'deepseek-ai/DeepSeek-R1',
286
+ // Grok mappings
287
+ 'grok 4 1 fast': 'xai-org/grok-fast',
288
+ 'grok 4 fast': 'xai-org/grok-fast',
289
+ 'grok code fast 1': 'xai-org/grok-code',
290
+ 'grok 3 mini': 'xai-org/grok-mini',
291
+ 'grok 3 mini beta': 'xai-org/grok-mini',
292
+ 'grok 4 20 multi agent beta': 'xai-org/grok-4',
293
+ 'grok 4 20 beta': 'xai-org/grok-4',
294
+ 'grok 4': 'xai-org/grok-4',
295
+ 'grok 3': 'xai-org/grok-3',
296
+ 'grok 3 beta': 'xai-org/grok-3',
297
+ 'grok 2 1212': 'xai-org/grok-2',
298
+ 'xai grok 4 1 fast': 'xai-org/grok-fast',
299
+ 'xai grok 4 fast': 'xai-org/grok-fast',
300
+ 'xai grok 3 mini': 'xai-org/grok-mini',
301
+ 'xai grok 3': 'xai-org/grok-3',
302
+ 'xai grok 4': 'xai-org/grok-4',
303
+ 'xai grok 2 1212': 'xai-org/grok-2',
304
+ // GLM mappings
305
+ 'glm 4 6v': 'THUDM/glm-4v-9b',
306
+ 'glm 5 turbo': 'THUDM/glm-5-turbo',
307
+ // MiniMax mappings
308
+ 'minimax m2 7': 'MiniMax/MiniMax-M2.7',
309
+ 'minimax 01': 'MiniMax/MiniMax-Text-01',
310
+ 'minimax m2 her': 'MiniMax/MiniMax-M2',
311
+ 'minimax m1': 'MiniMax/MiniMax-M1',
312
+ // Phi mappings
313
+ 'phi 4': 'microsoft/phi-4',
314
  // FLUX detailed mappings
315
  'flux 1 dev': 'black-forest-labs/FLUX.1-dev',
316
  'flux dev': 'black-forest-labs/FLUX.1-dev',
 
342
  'black-forest-labs/FLUX.2-klein-9B': 9,
343
  // Mistral family
344
  'mistralai/Mistral-Large-Instruct-2407': 123,
345
+ 'mistralai/Mistral-Large-Instruct-2411': 675,
346
+ // Qwen family
347
+ 'Alibaba/Qwen-Turbo': 14,
348
+ 'Qwen/Qwen2.5-Coder-7B-Instruct': 7,
349
+ 'Qwen/Qwen2.5-Coder-32B-Instruct': 32,
350
+ 'Qwen/Qwen2.5-7B-Instruct': 7,
351
+ 'Qwen/Qwen2-VL-7B-Instruct': 7,
352
+ 'Qwen/Qwen2-VL-72B-Instruct': 72,
353
+ // DeepSeek family
354
+ 'deepseek-ai/DeepSeek-V3': 671,
355
+ 'deepseek-ai/DeepSeek-R1': 671,
356
+ // Microsoft
357
+ 'microsoft/phi-4': 14,
358
+ // MiniMax
359
+ 'MiniMax/MiniMax-M2.7': 230,
360
  };
361
 
362
  // Propagate capabilities and size from benchmarks, OpenRouter, or HF Hub to all other providers' models.
 
398
  const n = normName(model.name);
399
 
400
  // 0. MANUAL OVERRIDE: Link common models to their HF IDs
401
+ if (!model.hf_id) {
402
+ for (const [key, val] of Object.entries(MANUAL_HF_ID_MAP)) {
403
+ if (n === key || n.endsWith(' ' + key) || n.endsWith('/' + key)) {
404
+ model.hf_id = val; break;
405
+ }
406
+ }
407
  }
408
 
409
+ // Apply size from manual map if available (even if model already has size, manual is high confidence)
410
+ if (model.hf_id && MANUAL_SIZE_MAP[model.hf_id]) {
411
+ if (model.size_b !== MANUAL_SIZE_MAP[model.hf_id]) {
 
412
  model.size_b = MANUAL_SIZE_MAP[model.hf_id];
413
  propagatedSize++;
 
 
 
 
414
  }
415
+ } else if (model.hf_id && !model.size_b) {
416
+ // Fallback to benchmarks if size missing
417
+ const size = hfIdToSize.get(model.hf_id.toLowerCase());
418
+ if (size) { model.size_b = size; propagatedSize++; }
419
  }
420
 
421
 
 
468
  }
469
 
470
  // 6. QUEUE: Still missing size? Try Hub API or Ollama
471
+ // Skip models that we've previously marked as private/unauthorized
472
+ if (!model.size_b && !model.hf_private && (model.name.includes('/') || model.hf_id)) {
473
+ hfLookupQueue.push(model);
474
+ } else if (!model.size_b && model.type === 'chat') {
475
+ ollamaLookupQueue.push(model);
 
476
  }
477
  }
478
  }
src/App.tsx CHANGED
@@ -18,6 +18,7 @@ interface Model {
18
  display_name?: string
19
  hf_id?: string
20
  ollama_id?: string
 
21
  }
22
 
23
  interface Provider {
@@ -547,6 +548,9 @@ function App() {
547
  </a>
548
  </div>
549
  )}
 
 
 
550
  {model.ollama_id && (
551
  <div className="tooltip-row">
552
  <strong>Ollama:</strong>
 
18
  display_name?: string
19
  hf_id?: string
20
  ollama_id?: string
21
+ hf_private?: boolean
22
  }
23
 
24
  interface Provider {
 
548
  </a>
549
  </div>
550
  )}
551
+ {!model.hf_id && model.hf_private && (
552
+ <div className="tooltip-row"><strong>HF:</strong> Proprietary API</div>
553
+ )}
554
  {model.ollama_id && (
555
  <div className="tooltip-row">
556
  <strong>Ollama:</strong>