CrispStrobe commited on
Commit
db007de
Β·
1 Parent(s): cf1221a

feat: integrate Ollama Registry, vLLM-style param estimation, and enhanced model info tooltips

Browse files
Files changed (4) hide show
  1. data/providers.json +620 -614
  2. scripts/fetch-providers.js +100 -14
  3. src/App.css +4 -2
  4. src/App.tsx +20 -0
data/providers.json CHANGED
@@ -998,22 +998,22 @@
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",
@@ -1062,6 +1062,18 @@
1062
  "context_window": 200000,
1063
  "hf_private": true
1064
  },
 
 
 
 
 
 
 
 
 
 
 
 
1065
  {
1066
  "name": "openai-responses/gpt-5-nano",
1067
  "type": "vision",
@@ -1106,18 +1118,6 @@
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",
@@ -1128,22 +1128,22 @@
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",
@@ -1155,20 +1155,19 @@
1155
  "size_b": 13
1156
  },
1157
  {
1158
- "name": "google/gemini-2.5-flash-lite",
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
- "reasoning"
1166
  ],
1167
- "context_window": 1048576,
1168
  "hf_private": true
1169
  },
1170
  {
1171
- "name": "google/gemini-2.0-flash-001",
1172
  "type": "vision",
1173
  "input_price_per_1m": 0.1,
1174
  "output_price_per_1m": 0.4,
@@ -1176,32 +1175,31 @@
1176
  "capabilities": [
1177
  "vision"
1178
  ],
1179
- "context_window": 1048576,
1180
  "hf_private": true
1181
  },
1182
  {
1183
- "name": "azure/openai-responses/gpt-4.1-nano",
1184
- "type": "vision",
1185
  "input_price_per_1m": 0.1,
1186
- "output_price_per_1m": 0.4,
1187
  "currency": "USD",
1188
- "capabilities": [
1189
- "vision"
1190
- ],
1191
- "context_window": 1047576,
1192
- "hf_private": true
1193
  },
1194
  {
1195
- "name": "azure/gpt-4.1-nano",
1196
- "type": "vision",
1197
  "input_price_per_1m": 0.1,
1198
- "output_price_per_1m": 0.4,
1199
  "currency": "USD",
 
 
1200
  "capabilities": [
1201
- "vision"
 
1202
  ],
1203
- "context_window": 1047576,
1204
- "hf_private": true
1205
  },
1206
  {
1207
  "name": "openai-responses/gpt-4.1-nano",
@@ -1241,27 +1239,29 @@
1241
  "hf_private": true
1242
  },
1243
  {
1244
- "name": "together/meta-llama/Meta-Llama-3-8B-Instruct-Lite",
1245
- "type": "chat",
1246
  "input_price_per_1m": 0.1,
1247
- "output_price_per_1m": 0.1,
1248
  "currency": "USD",
1249
- "context_window": 8192,
1250
- "size_b": 8
 
 
 
 
1251
  },
1252
  {
1253
- "name": "groq/openai/gpt-oss-20b",
1254
- "type": "chat",
1255
  "input_price_per_1m": 0.1,
1256
- "output_price_per_1m": 0.5,
1257
  "currency": "USD",
1258
- "context_window": 131072,
1259
- "size_b": 20,
1260
  "capabilities": [
1261
- "tools",
1262
- "reasoning"
1263
  ],
1264
- "hf_id": "openai/gpt-oss-20b"
 
1265
  },
1266
  {
1267
  "name": "mistral/devstral-small-latest",
@@ -1299,18 +1299,18 @@
1299
  "size_b": 24
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",
@@ -1396,17 +1396,16 @@
1396
  "hf_id": "deepseek-ai/DeepSeek-R1"
1397
  },
1398
  {
1399
- "name": "groq/openai/gpt-oss-120b",
1400
  "type": "chat",
1401
  "input_price_per_1m": 0.15,
1402
- "output_price_per_1m": 0.75,
1403
  "currency": "USD",
1404
- "context_window": 131072,
1405
- "size_b": 120,
1406
  "capabilities": [
1407
- "tools",
1408
  "reasoning"
1409
  ],
 
 
1410
  "hf_id": "openai/gpt-oss-120b"
1411
  },
1412
  {
@@ -1437,16 +1436,17 @@
1437
  "hf_id": "mistralai/Mistral-Small-4-119B-2603"
1438
  },
1439
  {
1440
- "name": "nebius/openai/gpt-oss-120b",
1441
  "type": "chat",
1442
  "input_price_per_1m": 0.15,
1443
- "output_price_per_1m": 0.6,
1444
  "currency": "USD",
 
 
1445
  "capabilities": [
 
1446
  "reasoning"
1447
  ],
1448
- "context_window": 131000,
1449
- "size_b": 120,
1450
  "hf_id": "openai/gpt-oss-120b"
1451
  },
1452
  {
@@ -1472,80 +1472,66 @@
1472
  "size_b": 8
1473
  },
1474
  {
1475
- "name": "openai-responses/gpt-5.4-nano",
1476
- "type": "vision",
1477
- "input_price_per_1m": 0.2,
1478
- "output_price_per_1m": 1.25,
1479
- "currency": "USD",
1480
- "capabilities": [
1481
- "vision",
1482
- "reasoning"
1483
- ],
1484
- "context_window": 400000,
1485
- "hf_private": true
1486
- },
1487
- {
1488
- "name": "xai/grok-4-1-fast-reasoning",
1489
- "type": "vision",
1490
  "input_price_per_1m": 0.2,
1491
- "output_price_per_1m": 0.5,
1492
  "currency": "USD",
1493
- "capabilities": [
1494
- "vision"
1495
- ],
1496
- "context_window": 2000000,
1497
- "hf_private": true
1498
  },
1499
  {
1500
- "name": "xai/grok-4-fast-non-reasoning",
1501
- "type": "vision",
1502
  "input_price_per_1m": 0.2,
1503
- "output_price_per_1m": 0.5,
1504
  "currency": "USD",
1505
- "context_window": 2000000,
 
1506
  "capabilities": [
1507
- "vision",
1508
  "tools",
1509
  "reasoning"
1510
  ],
1511
- "hf_private": true
1512
  },
1513
  {
1514
- "name": "xai/grok-4-fast",
1515
- "type": "vision",
1516
  "input_price_per_1m": 0.2,
1517
- "output_price_per_1m": 0.5,
1518
  "currency": "USD",
1519
- "context_window": 2000000,
1520
  "capabilities": [
1521
- "vision",
1522
  "tools",
1523
  "reasoning"
1524
  ],
1525
- "hf_private": true
 
1526
  },
1527
  {
1528
- "name": "xai/grok-4-1-fast-non-reasoning",
1529
  "type": "vision",
1530
  "input_price_per_1m": 0.2,
1531
- "output_price_per_1m": 0.5,
1532
  "currency": "USD",
1533
  "capabilities": [
1534
  "vision"
1535
  ],
1536
- "context_window": 2000000,
1537
- "hf_private": true
 
1538
  },
1539
  {
1540
- "name": "xai/grok-code-fast-1",
1541
  "type": "vision",
1542
  "input_price_per_1m": 0.2,
1543
- "output_price_per_1m": 1.5,
1544
  "currency": "USD",
1545
  "capabilities": [
1546
- "vision"
 
1547
  ],
1548
- "context_window": 256000,
1549
  "hf_private": true
1550
  },
1551
  {
@@ -1590,63 +1576,68 @@
1590
  "hf_id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
1591
  },
1592
  {
1593
- "name": "together/meta-llama/LlamaGuard-2-8b",
1594
- "type": "chat",
1595
- "input_price_per_1m": 0.2,
1596
- "output_price_per_1m": 0.2,
1597
- "currency": "USD",
1598
- "context_window": 8192,
1599
- "size_b": 8
1600
- },
1601
- {
1602
- "name": "deepinfra/zai-org/GLM-4.5-Air",
1603
- "type": "chat",
1604
  "input_price_per_1m": 0.2,
1605
- "output_price_per_1m": 1.1,
1606
  "currency": "USD",
1607
- "context_window": 131072,
1608
  "capabilities": [
1609
- "tools",
1610
- "reasoning"
1611
  ],
1612
- "hf_id": "zai-org/GLM-4.5-Air",
1613
- "size_b": 110.5
1614
  },
1615
  {
1616
- "name": "deepinfra/Qwen/Qwen3-235B-A22B",
1617
- "type": "chat",
1618
  "input_price_per_1m": 0.2,
1619
- "output_price_per_1m": 0.6,
1620
  "currency": "USD",
1621
- "context_window": 40000,
1622
- "size_b": 235,
1623
  "capabilities": [
 
1624
  "tools",
1625
  "reasoning"
1626
  ],
1627
- "hf_id": "Qwen/Qwen3-235B-A22B"
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/Meta-Llama-3.1-70B-Instruct",
1644
- "type": "chat",
1645
- "input_price_per_1m": 0.23,
1646
- "output_price_per_1m": 0.4,
1647
  "currency": "USD",
1648
- "context_window": 130815,
1649
- "size_b": 70
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1650
  },
1651
  {
1652
  "name": "deepinfra/Qwen/Qwen2.5-72B-Instruct",
@@ -1674,6 +1665,15 @@
1674
  ],
1675
  "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
1676
  },
 
 
 
 
 
 
 
 
 
1677
  {
1678
  "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct",
1679
  "type": "chat",
@@ -1687,18 +1687,6 @@
1687
  ],
1688
  "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
1689
  },
1690
- {
1691
- "name": "google/gemini-3.1-flash-lite-preview",
1692
- "type": "vision",
1693
- "input_price_per_1m": 0.25,
1694
- "output_price_per_1m": 1.5,
1695
- "currency": "USD",
1696
- "capabilities": [
1697
- "vision"
1698
- ],
1699
- "context_window": 1048576,
1700
- "hf_private": true
1701
- },
1702
  {
1703
  "name": "azure/gpt-5-mini",
1704
  "type": "vision",
@@ -1736,6 +1724,18 @@
1736
  "context_window": 1048576,
1737
  "hf_private": true
1738
  },
 
 
 
 
 
 
 
 
 
 
 
 
1739
  {
1740
  "name": "mistral/open-mistral-7b",
1741
  "type": "chat",
@@ -1780,79 +1780,47 @@
1780
  "hf_private": true
1781
  },
1782
  {
1783
- "name": "google/gemini-2.5-flash",
1784
- "type": "vision",
1785
- "input_price_per_1m": 0.3,
1786
- "output_price_per_1m": 2.5,
1787
- "currency": "USD",
1788
- "capabilities": [
1789
- "vision",
1790
- "reasoning"
1791
- ],
1792
- "context_window": 1048576,
1793
- "hf_private": true
1794
- },
1795
- {
1796
- "name": "parasail/parasail-gemma3-27b-it",
1797
  "type": "chat",
1798
  "input_price_per_1m": 0.3,
1799
- "output_price_per_1m": 0.5,
1800
- "currency": "USD",
1801
- "context_window": 128000,
1802
- "size_b": 27
1803
- },
1804
- {
1805
- "name": "minimaxi/MiniMax-M2.7",
1806
- "type": "vision",
1807
- "input_price_per_1m": 0.3,
1808
- "output_price_per_1m": 1.2,
1809
  "currency": "USD",
1810
- "capabilities": [
1811
- "vision",
1812
- "reasoning"
1813
- ],
1814
- "context_window": 200000,
1815
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
1816
- "hf_private": true
1817
  },
1818
  {
1819
- "name": "minimaxi/MiniMax-M2",
1820
  "type": "chat",
1821
  "input_price_per_1m": 0.3,
1822
- "output_price_per_1m": 1.2,
1823
  "currency": "USD",
1824
- "context_window": 200000,
1825
  "capabilities": [
1826
  "tools",
1827
  "reasoning"
1828
  ],
1829
- "hf_id": "MiniMaxAI/MiniMax-M2",
1830
- "size_b": 228.7
1831
  },
1832
  {
1833
- "name": "minimaxi/MiniMax-M2.5",
1834
- "type": "vision",
1835
  "input_price_per_1m": 0.3,
1836
- "output_price_per_1m": 1.2,
1837
  "currency": "USD",
1838
- "capabilities": [
1839
- "vision",
1840
- "reasoning"
1841
- ],
1842
- "context_window": 200000,
1843
- "hf_id": "MiniMaxAI/MiniMax-M2.5",
1844
- "size_b": 228.7
1845
  },
1846
  {
1847
- "name": "xai/grok-3-mini",
1848
- "type": "chat",
1849
  "input_price_per_1m": 0.3,
1850
- "output_price_per_1m": 0.5,
1851
  "currency": "USD",
1852
  "capabilities": [
1853
- "reasoning"
1854
  ],
1855
- "context_window": 131072,
1856
  "hf_private": true
1857
  },
1858
  {
@@ -1882,7 +1850,7 @@
1882
  "hf_private": true
1883
  },
1884
  {
1885
- "name": "novita/MiniMax-M2.7",
1886
  "type": "vision",
1887
  "input_price_per_1m": 0.3,
1888
  "output_price_per_1m": 1.2,
@@ -1896,7 +1864,20 @@
1896
  "hf_private": true
1897
  },
1898
  {
1899
- "name": "novita/minimax-m2.7",
 
 
 
 
 
 
 
 
 
 
 
 
 
1900
  "type": "vision",
1901
  "input_price_per_1m": 0.3,
1902
  "output_price_per_1m": 1.2,
@@ -1910,35 +1891,42 @@
1910
  "hf_private": true
1911
  },
1912
  {
1913
- "name": "novita/deepseek/deepseek-r1-distill-qwen-32b",
1914
  "type": "chat",
1915
  "input_price_per_1m": 0.3,
1916
- "output_price_per_1m": 0.3,
1917
  "currency": "USD",
1918
- "context_window": 12800,
1919
- "size_b": 32,
1920
  "capabilities": [
1921
  "reasoning"
1922
  ],
1923
- "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
 
1924
  },
1925
  {
1926
- "name": "together/Qwen/Qwen2.5-7B-Instruct-Turbo",
1927
- "type": "chat",
1928
  "input_price_per_1m": 0.3,
1929
- "output_price_per_1m": 0.3,
1930
  "currency": "USD",
1931
- "context_window": 32768,
1932
- "size_b": 7
 
 
 
 
1933
  },
1934
  {
1935
- "name": "mistral/codestral-latest",
1936
  "type": "chat",
1937
  "input_price_per_1m": 0.3,
1938
- "output_price_per_1m": 0.9,
1939
  "currency": "USD",
1940
- "context_window": 131072,
1941
- "hf_private": true
 
 
 
 
1942
  },
1943
  {
1944
  "name": "nebius/deepseek-ai/DeepSeek-V3.2",
@@ -1954,57 +1942,69 @@
1954
  "size_b": 685.4
1955
  },
1956
  {
1957
- "name": "nebius/minimax/minimax-m2.1",
1958
- "type": "chat",
1959
  "input_price_per_1m": 0.3,
1960
  "output_price_per_1m": 1.2,
1961
  "currency": "USD",
1962
  "capabilities": [
 
1963
  "reasoning"
1964
  ],
1965
  "context_window": 200000,
1966
- "hf_id": "MiniMaxAI/MiniMax-M2.1",
1967
- "size_b": 228.7
1968
  },
1969
  {
1970
- "name": "coding/gemini-2.5-flash",
1971
- "type": "vision",
1972
  "input_price_per_1m": 0.3,
1973
- "output_price_per_1m": 2.5,
1974
  "currency": "USD",
 
1975
  "capabilities": [
1976
- "vision",
1977
  "reasoning"
1978
  ],
1979
- "context_window": 1048576,
1980
- "hf_private": true
1981
  },
1982
  {
1983
- "name": "deepinfra/deepseek-ai/DeepSeek-V3.1",
1984
- "type": "chat",
1985
  "input_price_per_1m": 0.3,
1986
- "output_price_per_1m": 1,
1987
  "currency": "USD",
1988
- "context_window": 163840,
1989
  "capabilities": [
1990
- "tools",
1991
  "reasoning"
1992
  ],
1993
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
1994
- "size_b": 671
 
1995
  },
1996
  {
1997
- "name": "alibaba/qwen3-coder-flash",
1998
  "type": "vision",
1999
  "input_price_per_1m": 0.3,
2000
- "output_price_per_1m": 1.5,
2001
  "currency": "USD",
2002
  "capabilities": [
2003
- "vision"
 
2004
  ],
2005
  "context_window": 1048576,
2006
  "hf_private": true
2007
  },
 
 
 
 
 
 
 
 
 
2008
  {
2009
  "name": "deepinfra/meta-llama/Llama-3.2-90B-Vision-Instruct",
2010
  "type": "chat",
@@ -2064,6 +2064,31 @@
2064
  "context_window": 1047576,
2065
  "hf_private": true
2066
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2067
  {
2068
  "name": "openai-responses/gpt-4.1-mini",
2069
  "type": "vision",
@@ -2112,34 +2137,17 @@
2112
  "hf_private": true
2113
  },
2114
  {
2115
- "name": "mistral/devstral-latest",
2116
- "type": "chat",
2117
- "input_price_per_1m": 0.4,
2118
- "output_price_per_1m": 2,
2119
- "currency": "USD",
2120
- "context_window": 256000,
2121
- "hf_private": true
2122
- },
2123
- {
2124
- "name": "mistral/mistral-medium-latest",
2125
- "type": "chat",
2126
- "input_price_per_1m": 0.4,
2127
- "output_price_per_1m": 2,
2128
- "currency": "USD",
2129
- "context_window": 131072,
2130
- "hf_private": true
2131
- },
2132
- {
2133
- "name": "mistral/devstral-medium-2507",
2134
  "type": "chat",
2135
  "input_price_per_1m": 0.4,
2136
  "output_price_per_1m": 2,
2137
  "currency": "USD",
2138
- "context_window": 131072,
2139
  "capabilities": [
2140
- "tools"
2141
  ],
2142
- "hf_private": true
 
 
2143
  },
2144
  {
2145
  "name": "nebius/Qwen/Qwen3-Coder-480B-A35B-Instruct",
@@ -2155,45 +2163,37 @@
2155
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2156
  },
2157
  {
2158
- "name": "nebius/zai-org/GLM-4.7",
2159
  "type": "chat",
2160
  "input_price_per_1m": 0.4,
2161
  "output_price_per_1m": 2,
2162
  "currency": "USD",
2163
- "capabilities": [
2164
- "reasoning"
2165
- ],
2166
- "context_window": 200000,
2167
- "hf_id": "zai-org/GLM-4.7",
2168
- "size_b": 358.3
2169
  },
2170
  {
2171
- "name": "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct",
2172
  "type": "chat",
2173
  "input_price_per_1m": 0.4,
2174
- "output_price_per_1m": 1.6,
2175
  "currency": "USD",
2176
- "context_window": 262144,
2177
- "size_b": 480,
2178
  "capabilities": [
2179
  "tools"
2180
  ],
2181
- "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2182
  },
2183
  {
2184
- "name": "alibaba/qwen-plus",
2185
  "type": "chat",
2186
  "input_price_per_1m": 0.4,
2187
- "output_price_per_1m": 1.2,
2188
  "currency": "USD",
2189
- "context_window": 131072,
2190
- "capabilities": [
2191
- "tools"
2192
- ],
2193
  "hf_private": true
2194
  },
2195
  {
2196
- "name": "google/gemini-3.1-flash-image-preview",
2197
  "type": "vision",
2198
  "input_price_per_1m": 0.5,
2199
  "output_price_per_1m": 2,
@@ -2206,7 +2206,7 @@
2206
  "hf_private": true
2207
  },
2208
  {
2209
- "name": "google/gemini-3-flash-preview",
2210
  "type": "vision",
2211
  "input_price_per_1m": 0.5,
2212
  "output_price_per_1m": 3,
@@ -2219,83 +2219,83 @@
2219
  "hf_private": true
2220
  },
2221
  {
2222
- "name": "vertex/gemini-3-flash-preview",
2223
- "type": "vision",
2224
  "input_price_per_1m": 0.5,
2225
- "output_price_per_1m": 3,
2226
  "currency": "USD",
 
2227
  "capabilities": [
2228
- "vision",
2229
- "reasoning"
2230
  ],
2231
- "context_window": 1048576,
2232
- "hf_private": true
2233
  },
2234
  {
2235
- "name": "vertex/gemini-3.1-flash-image-preview",
2236
  "type": "vision",
2237
  "input_price_per_1m": 0.5,
2238
- "output_price_per_1m": 2,
2239
  "currency": "USD",
2240
  "capabilities": [
2241
  "vision",
2242
  "reasoning"
2243
  ],
2244
- "context_window": 131072,
2245
- "hf_private": true
 
2246
  },
2247
  {
2248
- "name": "mistral/mistral-large-latest",
2249
  "type": "chat",
2250
  "input_price_per_1m": 0.5,
2251
  "output_price_per_1m": 1.5,
2252
  "currency": "USD",
2253
- "context_window": 131072,
2254
  "capabilities": [
2255
- "tools"
 
2256
  ],
2257
- "hf_private": true
 
2258
  },
2259
  {
2260
- "name": "nebius/moonshotai/Kimi-K2-Instruct",
2261
- "type": "chat",
2262
  "input_price_per_1m": 0.5,
2263
- "output_price_per_1m": 2.4,
2264
  "currency": "USD",
2265
- "context_window": 131000,
2266
  "capabilities": [
2267
- "tools"
 
2268
  ],
2269
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2270
- "size_b": 1026.5
2271
  },
2272
  {
2273
- "name": "nebius/deepseek-ai/DeepSeek-V3-0324",
2274
- "type": "chat",
2275
  "input_price_per_1m": 0.5,
2276
- "output_price_per_1m": 1.5,
2277
  "currency": "USD",
2278
- "context_window": 128000,
2279
  "capabilities": [
2280
- "tools",
2281
  "reasoning"
2282
  ],
2283
- "size_b": 685,
2284
- "hf_id": "deepseek-ai/DeepSeek-V3-0324"
2285
  },
2286
  {
2287
- "name": "nebius/moonshotai/kimi-k2.5",
2288
- "type": "vision",
2289
  "input_price_per_1m": 0.5,
2290
- "output_price_per_1m": 2.5,
2291
  "currency": "USD",
 
2292
  "capabilities": [
2293
- "vision",
2294
- "reasoning"
2295
  ],
2296
- "context_window": 256000,
2297
- "hf_id": "moonshotai/Kimi-K2.5",
2298
- "size_b": 1058.6
2299
  },
2300
  {
2301
  "name": "novita/meta-llama/llama-3-70b-instruct",
@@ -2350,72 +2350,102 @@
2350
  "size_b": 1026.5
2351
  },
2352
  {
2353
- "name": "minimaxi/MiniMax-M2.7-highspeed",
2354
- "type": "vision",
2355
  "input_price_per_1m": 0.6,
2356
- "output_price_per_1m": 2.4,
2357
  "currency": "USD",
 
2358
  "capabilities": [
2359
- "vision",
2360
  "reasoning"
2361
  ],
2362
- "context_window": 200000,
2363
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
2364
- "hf_private": true
2365
  },
2366
  {
2367
- "name": "minimaxi/MiniMax-M2.5-highspeed",
2368
  "type": "vision",
2369
  "input_price_per_1m": 0.6,
2370
- "output_price_per_1m": 2.4,
2371
  "currency": "USD",
2372
  "capabilities": [
2373
  "vision",
2374
  "reasoning"
2375
  ],
2376
- "context_window": 200000,
2377
- "hf_id": "MiniMaxAI/MiniMax-M2.5",
2378
- "size_b": 228.7
2379
  },
2380
  {
2381
- "name": "moonshot/kimi-k2-thinking",
2382
  "type": "chat",
2383
  "input_price_per_1m": 0.6,
2384
- "output_price_per_1m": 2.5,
2385
  "currency": "USD",
2386
  "context_window": 131072,
2387
  "capabilities": [
2388
- "tools"
 
2389
  ],
2390
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2391
- "size_b": 1026.5
2392
  },
2393
  {
2394
- "name": "moonshot/kimi-k2-0711-preview",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2395
  "type": "chat",
2396
  "input_price_per_1m": 0.6,
2397
- "output_price_per_1m": 2.5,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2398
  "currency": "USD",
2399
- "context_window": 131072,
2400
  "capabilities": [
2401
- "tools"
 
2402
  ],
2403
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2404
- "size_b": 1026.5
 
2405
  },
2406
  {
2407
- "name": "moonshot/kimi-k2.5",
2408
  "type": "vision",
2409
  "input_price_per_1m": 0.6,
2410
- "output_price_per_1m": 3,
2411
  "currency": "USD",
2412
  "capabilities": [
2413
  "vision",
2414
  "reasoning"
2415
  ],
2416
- "context_window": 262144,
2417
- "hf_id": "moonshotai/Kimi-K2.5",
2418
- "size_b": 1058.6
2419
  },
2420
  {
2421
  "name": "moonshot/kimi-k2-thinking-turbo",
@@ -2445,32 +2475,44 @@
2445
  "size_b": 1026.5
2446
  },
2447
  {
2448
- "name": "novita/zai-org/glm-4.6",
2449
  "type": "chat",
2450
  "input_price_per_1m": 0.6,
2451
- "output_price_per_1m": 2.2,
2452
  "currency": "USD",
2453
- "context_window": 204800,
2454
  "capabilities": [
2455
- "tools",
2456
- "reasoning"
2457
  ],
2458
- "hf_id": "zai-org/GLM-4.6",
2459
- "size_b": 356.8
2460
  },
2461
  {
2462
- "name": "novita/zai-org/glm-4.5",
2463
  "type": "chat",
2464
  "input_price_per_1m": 0.6,
2465
- "output_price_per_1m": 2.2,
2466
  "currency": "USD",
2467
  "context_window": 131072,
2468
  "capabilities": [
2469
- "tools",
 
 
 
 
 
 
 
 
 
 
 
 
2470
  "reasoning"
2471
  ],
2472
- "hf_id": "zai-org/GLM-4.5",
2473
- "size_b": 358.3
 
2474
  },
2475
  {
2476
  "name": "zai/GLM-4.6",
@@ -2511,48 +2553,6 @@
2511
  "hf_id": "zai-org/GLM-4.5",
2512
  "size_b": 358.3
2513
  },
2514
- {
2515
- "name": "nebius/zai-org/GLM-4.5",
2516
- "type": "chat",
2517
- "input_price_per_1m": 0.6,
2518
- "output_price_per_1m": 2.2,
2519
- "currency": "USD",
2520
- "context_window": 128000,
2521
- "capabilities": [
2522
- "tools",
2523
- "reasoning"
2524
- ],
2525
- "hf_id": "zai-org/GLM-4.5",
2526
- "size_b": 358.3
2527
- },
2528
- {
2529
- "name": "deepinfra/zai-org/GLM-4.5",
2530
- "type": "chat",
2531
- "input_price_per_1m": 0.6,
2532
- "output_price_per_1m": 2.2,
2533
- "currency": "USD",
2534
- "context_window": 131072,
2535
- "capabilities": [
2536
- "tools",
2537
- "reasoning"
2538
- ],
2539
- "hf_id": "zai-org/GLM-4.5",
2540
- "size_b": 358.3
2541
- },
2542
- {
2543
- "name": "alibaba/qwen3.5",
2544
- "type": "vision",
2545
- "input_price_per_1m": 0.6,
2546
- "output_price_per_1m": 3.6,
2547
- "currency": "USD",
2548
- "capabilities": [
2549
- "vision",
2550
- "reasoning"
2551
- ],
2552
- "context_window": 256000,
2553
- "size_b": 9,
2554
- "hf_id": "Qwen/Qwen3.5-9B"
2555
- },
2556
  {
2557
  "name": "novita/microsoft/wizardlm-2-8x22b",
2558
  "type": "chat",
@@ -2636,6 +2636,15 @@
2636
  "context_window": 400000,
2637
  "hf_private": true
2638
  },
 
 
 
 
 
 
 
 
 
2639
  {
2640
  "name": "novita/deepseek/deepseek-r1-distill-llama-70b",
2641
  "type": "chat",
@@ -2676,15 +2685,6 @@
2676
  "hf_id": "deepseek-ai/DeepSeek-R1-0528",
2677
  "size_b": 684.5
2678
  },
2679
- {
2680
- "name": "deepinfra/meta-llama/Meta-Llama-3.1-405B-Instruct",
2681
- "type": "chat",
2682
- "input_price_per_1m": 0.8,
2683
- "output_price_per_1m": 0.8,
2684
- "currency": "USD",
2685
- "context_window": 130815,
2686
- "size_b": 405
2687
- },
2688
  {
2689
  "name": "deepinfra/deepseek-ai/DeepSeek-R1",
2690
  "type": "chat",
@@ -2771,7 +2771,7 @@
2771
  "hf_private": true
2772
  },
2773
  {
2774
- "name": "bedrock/claude-haiku-4-5",
2775
  "type": "vision",
2776
  "input_price_per_1m": 1,
2777
  "output_price_per_1m": 5,
@@ -2783,7 +2783,7 @@
2783
  "hf_private": true
2784
  },
2785
  {
2786
- "name": "anthropic/claude-haiku-4-5",
2787
  "type": "vision",
2788
  "input_price_per_1m": 1,
2789
  "output_price_per_1m": 5,
@@ -2791,7 +2791,7 @@
2791
  "capabilities": [
2792
  "vision"
2793
  ],
2794
- "context_window": 200000,
2795
  "hf_private": true
2796
  },
2797
  {
@@ -2844,6 +2844,18 @@
2844
  "hf_id": "zai-org/GLM-5",
2845
  "size_b": 753.9
2846
  },
 
 
 
 
 
 
 
 
 
 
 
 
2847
  {
2848
  "name": "zai/GLM-5",
2849
  "type": "vision",
@@ -2884,18 +2896,6 @@
2884
  "hf_id": "moonshotai/Kimi-K2-Instruct",
2885
  "size_b": 1026.5
2886
  },
2887
- {
2888
- "name": "alibaba/qwen3-coder-plus",
2889
- "type": "vision",
2890
- "input_price_per_1m": 1,
2891
- "output_price_per_1m": 5,
2892
- "currency": "USD",
2893
- "capabilities": [
2894
- "vision"
2895
- ],
2896
- "context_window": 1048576,
2897
- "hf_private": true
2898
- },
2899
  {
2900
  "name": "azure/o4-mini",
2901
  "type": "vision",
@@ -2944,6 +2944,15 @@
2944
  "context_window": 200000,
2945
  "hf_private": true
2946
  },
 
 
 
 
 
 
 
 
 
2947
  {
2948
  "name": "moonshot/kimi-k2-turbo-preview",
2949
  "type": "chat",
@@ -2958,25 +2967,15 @@
2958
  "size_b": 1026.5
2959
  },
2960
  {
2961
- "name": "together/Qwen/Qwen2.5-72B-Instruct-Turbo",
2962
- "type": "chat",
2963
- "input_price_per_1m": 1.2,
2964
- "output_price_per_1m": 1.2,
2965
- "currency": "USD",
2966
- "context_window": 32768,
2967
- "size_b": 72
2968
- },
2969
- {
2970
- "name": "google/gemini-2.5-pro",
2971
  "type": "vision",
2972
  "input_price_per_1m": 1.25,
2973
  "output_price_per_1m": 10,
2974
  "currency": "USD",
2975
  "capabilities": [
2976
- "vision",
2977
  "reasoning"
2978
  ],
2979
- "context_window": 1048576,
2980
  "hf_private": true
2981
  },
2982
  {
@@ -2992,16 +2991,18 @@
2992
  "hf_private": true
2993
  },
2994
  {
2995
- "name": "azure/gpt-5",
2996
- "type": "vision",
2997
  "input_price_per_1m": 1.25,
2998
- "output_price_per_1m": 10,
2999
  "currency": "USD",
 
3000
  "capabilities": [
 
3001
  "reasoning"
3002
  ],
3003
- "context_window": 200000,
3004
- "hf_private": true
3005
  },
3006
  {
3007
  "name": "openai-responses/gpt-5.1",
@@ -3030,7 +3031,7 @@
3030
  "hf_private": true
3031
  },
3032
  {
3033
- "name": "openai-responses/gpt-5",
3034
  "type": "vision",
3035
  "input_price_per_1m": 1.25,
3036
  "output_price_per_1m": 10,
@@ -3043,7 +3044,7 @@
3043
  "hf_private": true
3044
  },
3045
  {
3046
- "name": "openai-responses/gpt-5.1-codex",
3047
  "type": "vision",
3048
  "input_price_per_1m": 1.25,
3049
  "output_price_per_1m": 10,
@@ -3108,21 +3109,20 @@
3108
  "hf_private": true
3109
  },
3110
  {
3111
- "name": "together/deepseek-ai/DeepSeek-V3",
3112
- "type": "chat",
3113
  "input_price_per_1m": 1.25,
3114
- "output_price_per_1m": 1.25,
3115
  "currency": "USD",
3116
- "context_window": 131072,
3117
  "capabilities": [
3118
- "tools",
3119
  "reasoning"
3120
  ],
3121
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
3122
- "size_b": 671
3123
  },
3124
  {
3125
- "name": "coding/gemini-2.5-pro",
3126
  "type": "vision",
3127
  "input_price_per_1m": 1.25,
3128
  "output_price_per_1m": 10,
@@ -3164,18 +3164,6 @@
3164
  ],
3165
  "hf_private": true
3166
  },
3167
- {
3168
- "name": "azure/gpt-5.2",
3169
- "type": "vision",
3170
- "input_price_per_1m": 1.75,
3171
- "output_price_per_1m": 14,
3172
- "currency": "USD",
3173
- "capabilities": [
3174
- "reasoning"
3175
- ],
3176
- "context_window": 200000,
3177
- "hf_private": true
3178
- },
3179
  {
3180
  "name": "azure/openai-responses/gpt-5.2-codex",
3181
  "type": "vision",
@@ -3203,20 +3191,19 @@
3203
  "hf_private": true
3204
  },
3205
  {
3206
- "name": "openai-responses/gpt-5.2-codex",
3207
  "type": "vision",
3208
  "input_price_per_1m": 1.75,
3209
  "output_price_per_1m": 14,
3210
  "currency": "USD",
3211
  "capabilities": [
3212
- "vision",
3213
  "reasoning"
3214
  ],
3215
- "context_window": 400000,
3216
  "hf_private": true
3217
  },
3218
  {
3219
- "name": "openai-responses/gpt-5.3-codex",
3220
  "type": "vision",
3221
  "input_price_per_1m": 1.75,
3222
  "output_price_per_1m": 14,
@@ -3225,11 +3212,11 @@
3225
  "vision",
3226
  "reasoning"
3227
  ],
3228
- "context_window": 400000,
3229
  "hf_private": true
3230
  },
3231
  {
3232
- "name": "openai-responses/gpt-5.3-chat",
3233
  "type": "vision",
3234
  "input_price_per_1m": 1.75,
3235
  "output_price_per_1m": 14,
@@ -3238,7 +3225,7 @@
3238
  "vision",
3239
  "reasoning"
3240
  ],
3241
- "context_window": 128000,
3242
  "hf_private": true
3243
  },
3244
  {
@@ -3255,7 +3242,7 @@
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 +3255,7 @@
3268
  "hf_private": true
3269
  },
3270
  {
3271
- "name": "openai/gpt-5.2-chat",
3272
  "type": "vision",
3273
  "input_price_per_1m": 1.75,
3274
  "output_price_per_1m": 14,
@@ -3277,7 +3264,7 @@
3277
  "vision",
3278
  "reasoning"
3279
  ],
3280
- "context_window": 128000,
3281
  "hf_private": true
3282
  },
3283
  {
@@ -3294,33 +3281,20 @@
3294
  "hf_private": true
3295
  },
3296
  {
3297
- "name": "google/gemini-3-pro-preview",
3298
- "type": "vision",
3299
- "input_price_per_1m": 2,
3300
- "output_price_per_1m": 12,
3301
- "currency": "USD",
3302
- "capabilities": [
3303
- "vision",
3304
- "reasoning"
3305
- ],
3306
- "context_window": 1048576,
3307
- "hf_private": true
3308
- },
3309
- {
3310
- "name": "google/gemini-3.1-pro-preview",
3311
  "type": "vision",
3312
- "input_price_per_1m": 2,
3313
- "output_price_per_1m": 12,
3314
  "currency": "USD",
3315
  "capabilities": [
3316
  "vision",
3317
  "reasoning"
3318
  ],
3319
- "context_window": 1048576,
3320
  "hf_private": true
3321
  },
3322
  {
3323
- "name": "azure/openai-responses/gpt-4.1",
3324
  "type": "vision",
3325
  "input_price_per_1m": 2,
3326
  "output_price_per_1m": 8,
@@ -3332,7 +3306,7 @@
3332
  "hf_private": true
3333
  },
3334
  {
3335
- "name": "azure/gpt-4.1",
3336
  "type": "vision",
3337
  "input_price_per_1m": 2,
3338
  "output_price_per_1m": 8,
@@ -3369,25 +3343,16 @@
3369
  "hf_private": true
3370
  },
3371
  {
3372
- "name": "xai/grok-4.2-beta",
3373
  "type": "vision",
3374
  "input_price_per_1m": 2,
3375
- "output_price_per_1m": 6,
3376
  "currency": "USD",
3377
  "capabilities": [
3378
  "vision",
3379
  "reasoning"
3380
  ],
3381
- "context_window": 2000000,
3382
- "hf_private": true
3383
- },
3384
- {
3385
- "name": "xai/grok-2-1212",
3386
- "type": "chat",
3387
- "input_price_per_1m": 2,
3388
- "output_price_per_1m": 10,
3389
- "currency": "USD",
3390
- "context_window": 131072,
3391
  "hf_private": true
3392
  },
3393
  {
@@ -3404,7 +3369,7 @@
3404
  "hf_private": true
3405
  },
3406
  {
3407
- "name": "vertex/gemini-3-pro-preview",
3408
  "type": "vision",
3409
  "input_price_per_1m": 2,
3410
  "output_price_per_1m": 12,
@@ -3417,16 +3382,15 @@
3417
  "hf_private": true
3418
  },
3419
  {
3420
- "name": "vertex/gemini-3.1-pro-preview",
3421
  "type": "vision",
3422
  "input_price_per_1m": 2,
3423
- "output_price_per_1m": 12,
3424
  "currency": "USD",
3425
  "capabilities": [
3426
- "vision",
3427
- "reasoning"
3428
  ],
3429
- "context_window": 1048576,
3430
  "hf_private": true
3431
  },
3432
  {
@@ -3443,22 +3407,23 @@
3443
  "hf_private": true
3444
  },
3445
  {
3446
- "name": "openai/gpt-4.1",
3447
  "type": "vision",
3448
  "input_price_per_1m": 2,
3449
- "output_price_per_1m": 8,
3450
  "currency": "USD",
3451
  "capabilities": [
3452
- "vision"
 
3453
  ],
3454
- "context_window": 1047576,
3455
  "hf_private": true
3456
  },
3457
  {
3458
- "name": "mistral/pixtral-large-latest",
3459
  "type": "chat",
3460
  "input_price_per_1m": 2,
3461
- "output_price_per_1m": 5,
3462
  "currency": "USD",
3463
  "context_window": 131072,
3464
  "hf_private": true
@@ -3472,6 +3437,41 @@
3472
  "context_window": 128000,
3473
  "hf_private": true
3474
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3475
  {
3476
  "name": "azure/openai-responses/gpt-5.4",
3477
  "type": "vision",
@@ -3499,7 +3499,7 @@
3499
  "hf_private": true
3500
  },
3501
  {
3502
- "name": "openai/gpt-4o",
3503
  "type": "vision",
3504
  "input_price_per_1m": 2.5,
3505
  "output_price_per_1m": 10,
@@ -3510,6 +3510,19 @@
3510
  "context_window": 128000,
3511
  "hf_private": true
3512
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
3513
  {
3514
  "name": "openai/gpt-4o-2024-05-13",
3515
  "type": "vision",
@@ -3523,7 +3536,7 @@
3523
  "hf_private": true
3524
  },
3525
  {
3526
- "name": "openai/gpt-4o-2024-08-06",
3527
  "type": "vision",
3528
  "input_price_per_1m": 2.5,
3529
  "output_price_per_1m": 10,
@@ -3535,7 +3548,7 @@
3535
  "hf_private": true
3536
  },
3537
  {
3538
- "name": "openai/gpt-4o-2024-11-20",
3539
  "type": "vision",
3540
  "input_price_per_1m": 2.5,
3541
  "output_price_per_1m": 10,
@@ -3547,20 +3560,20 @@
3547
  "hf_private": true
3548
  },
3549
  {
3550
- "name": "openai/gpt-5.4",
3551
  "type": "vision",
3552
- "input_price_per_1m": 2.5,
3553
  "output_price_per_1m": 15,
3554
  "currency": "USD",
3555
  "capabilities": [
3556
  "vision",
3557
  "reasoning"
3558
  ],
3559
- "context_window": 1050000,
3560
  "hf_private": true
3561
  },
3562
  {
3563
- "name": "bedrock/claude-sonnet-4",
3564
  "type": "vision",
3565
  "input_price_per_1m": 3,
3566
  "output_price_per_1m": 15,
@@ -3569,11 +3582,11 @@
3569
  "vision",
3570
  "reasoning"
3571
  ],
3572
- "context_window": 200000,
3573
  "hf_private": true
3574
  },
3575
  {
3576
- "name": "bedrock/claude-sonnet-4-6",
3577
  "type": "vision",
3578
  "input_price_per_1m": 3,
3579
  "output_price_per_1m": 15,
@@ -3586,46 +3599,45 @@
3586
  "hf_private": true
3587
  },
3588
  {
3589
- "name": "bedrock/claude-3-7-sonnet",
3590
- "type": "vision",
3591
  "input_price_per_1m": 3,
3592
- "output_price_per_1m": 15,
3593
  "currency": "USD",
 
3594
  "capabilities": [
3595
- "vision",
3596
  "reasoning"
3597
  ],
3598
- "context_window": 200000,
3599
- "hf_private": true
3600
  },
3601
  {
3602
- "name": "bedrock/claude-sonnet-4-5",
3603
  "type": "vision",
3604
  "input_price_per_1m": 3,
3605
  "output_price_per_1m": 15,
3606
  "currency": "USD",
 
3607
  "capabilities": [
3608
- "vision",
3609
- "reasoning"
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,
@@ -3634,11 +3646,11 @@
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,
@@ -3647,19 +3659,20 @@
3647
  "vision",
3648
  "reasoning"
3649
  ],
3650
- "context_window": 1000000,
3651
  "hf_private": true
3652
  },
3653
  {
3654
- "name": "perplexity/sonar-pro",
3655
  "type": "vision",
3656
  "input_price_per_1m": 3,
3657
  "output_price_per_1m": 15,
3658
  "currency": "USD",
3659
- "context_window": 204800,
3660
  "capabilities": [
3661
- "vision"
 
3662
  ],
 
3663
  "hf_private": true
3664
  },
3665
  {
@@ -3675,7 +3688,7 @@
3675
  "hf_private": true
3676
  },
3677
  {
3678
- "name": "vertex/claude-3-7-sonnet",
3679
  "type": "vision",
3680
  "input_price_per_1m": 3,
3681
  "output_price_per_1m": 15,
@@ -3688,7 +3701,7 @@
3688
  "hf_private": true
3689
  },
3690
  {
3691
- "name": "vertex/claude-sonnet-4",
3692
  "type": "vision",
3693
  "input_price_per_1m": 3,
3694
  "output_price_per_1m": 15,
@@ -3701,19 +3714,20 @@
3701
  "hf_private": true
3702
  },
3703
  {
3704
- "name": "vertex/claude-3-5-sonnet",
3705
  "type": "vision",
3706
  "input_price_per_1m": 3,
3707
  "output_price_per_1m": 15,
3708
  "currency": "USD",
3709
  "capabilities": [
3710
- "vision"
 
3711
  ],
3712
  "context_window": 200000,
3713
  "hf_private": true
3714
  },
3715
  {
3716
- "name": "vertex/claude-sonnet-4-5",
3717
  "type": "vision",
3718
  "input_price_per_1m": 3,
3719
  "output_price_per_1m": 15,
@@ -3722,25 +3736,11 @@
3722
  "vision",
3723
  "reasoning"
3724
  ],
3725
- "context_window": 200000,
3726
  "hf_private": true
3727
  },
3728
  {
3729
- "name": "together/deepseek-ai/DeepSeek-R1",
3730
- "type": "chat",
3731
- "input_price_per_1m": 3,
3732
- "output_price_per_1m": 7,
3733
- "currency": "USD",
3734
- "context_window": 64000,
3735
- "capabilities": [
3736
- "tools",
3737
- "reasoning"
3738
- ],
3739
- "hf_id": "deepseek-ai/DeepSeek-R1",
3740
- "size_b": 684.5
3741
- },
3742
- {
3743
- "name": "coding/claude-sonnet-4-20250514",
3744
  "type": "vision",
3745
  "input_price_per_1m": 3,
3746
  "output_price_per_1m": 15,
@@ -3767,7 +3767,7 @@
3767
  "size_b": 684.5
3768
  },
3769
  {
3770
- "name": "bedrock/claude-opus-4-5",
3771
  "type": "vision",
3772
  "input_price_per_1m": 5,
3773
  "output_price_per_1m": 25,
@@ -3780,7 +3780,7 @@
3780
  "hf_private": true
3781
  },
3782
  {
3783
- "name": "bedrock/claude-opus-4-6",
3784
  "type": "vision",
3785
  "input_price_per_1m": 5,
3786
  "output_price_per_1m": 25,
@@ -3793,7 +3793,7 @@
3793
  "hf_private": true
3794
  },
3795
  {
3796
- "name": "anthropic/claude-opus-4-6",
3797
  "type": "vision",
3798
  "input_price_per_1m": 5,
3799
  "output_price_per_1m": 25,
@@ -3806,7 +3806,7 @@
3806
  "hf_private": true
3807
  },
3808
  {
3809
- "name": "anthropic/claude-opus-4-5",
3810
  "type": "vision",
3811
  "input_price_per_1m": 5,
3812
  "output_price_per_1m": 25,
@@ -3819,32 +3819,31 @@
3819
  "hf_private": true
3820
  },
3821
  {
3822
- "name": "xai/grok-3",
3823
- "type": "chat",
3824
  "input_price_per_1m": 5,
3825
- "output_price_per_1m": 25,
3826
  "currency": "USD",
3827
- "context_window": 131072,
3828
  "capabilities": [
3829
- "tools"
3830
  ],
 
3831
  "hf_private": true
3832
  },
3833
  {
3834
- "name": "vertex/claude-opus-4-6",
3835
- "type": "vision",
3836
  "input_price_per_1m": 5,
3837
  "output_price_per_1m": 25,
3838
  "currency": "USD",
 
3839
  "capabilities": [
3840
- "vision",
3841
- "reasoning"
3842
  ],
3843
- "context_window": 1000000,
3844
  "hf_private": true
3845
  },
3846
  {
3847
- "name": "vertex/claude-opus-4-5",
3848
  "type": "vision",
3849
  "input_price_per_1m": 5,
3850
  "output_price_per_1m": 25,
@@ -3857,15 +3856,16 @@
3857
  "hf_private": true
3858
  },
3859
  {
3860
- "name": "openai/chatgpt-4o",
3861
  "type": "vision",
3862
  "input_price_per_1m": 5,
3863
- "output_price_per_1m": 15,
3864
  "currency": "USD",
3865
  "capabilities": [
3866
- "vision"
 
3867
  ],
3868
- "context_window": 128000,
3869
  "hf_private": true
3870
  },
3871
  {
@@ -3882,7 +3882,7 @@
3882
  "hf_private": true
3883
  },
3884
  {
3885
- "name": "bedrock/claude-opus-4",
3886
  "type": "vision",
3887
  "input_price_per_1m": 15,
3888
  "output_price_per_1m": 75,
@@ -3908,33 +3908,33 @@
3908
  "hf_private": true
3909
  },
3910
  {
3911
- "name": "anthropic/claude-opus-4-1",
3912
  "type": "vision",
3913
  "input_price_per_1m": 15,
3914
- "output_price_per_1m": 75,
3915
  "currency": "USD",
3916
  "capabilities": [
3917
  "vision",
3918
  "reasoning"
3919
  ],
3920
- "context_window": 200000,
3921
  "hf_private": true
3922
  },
3923
  {
3924
- "name": "openai-responses/gpt-5-pro",
3925
  "type": "vision",
3926
  "input_price_per_1m": 15,
3927
- "output_price_per_1m": 120,
3928
  "currency": "USD",
3929
  "capabilities": [
3930
  "vision",
3931
  "reasoning"
3932
  ],
3933
- "context_window": 400000,
3934
  "hf_private": true
3935
  },
3936
  {
3937
- "name": "vertex/claude-opus-4-1",
3938
  "type": "vision",
3939
  "input_price_per_1m": 15,
3940
  "output_price_per_1m": 75,
@@ -3947,10 +3947,10 @@
3947
  "hf_private": true
3948
  },
3949
  {
3950
- "name": "vertex/claude-opus-4",
3951
  "type": "vision",
3952
  "input_price_per_1m": 15,
3953
- "output_price_per_1m": 75,
3954
  "currency": "USD",
3955
  "capabilities": [
3956
  "vision",
@@ -3960,10 +3960,10 @@
3960
  "hf_private": true
3961
  },
3962
  {
3963
- "name": "openai/o1",
3964
  "type": "vision",
3965
  "input_price_per_1m": 15,
3966
- "output_price_per_1m": 60,
3967
  "currency": "USD",
3968
  "capabilities": [
3969
  "vision",
@@ -3973,7 +3973,7 @@
3973
  "hf_private": true
3974
  },
3975
  {
3976
- "name": "coding/claude-opus-4-20250514",
3977
  "type": "vision",
3978
  "input_price_per_1m": 15,
3979
  "output_price_per_1m": 75,
@@ -4594,7 +4594,9 @@
4594
  "currency": "USD",
4595
  "capabilities": [
4596
  "image-gen"
4597
- ]
 
 
4598
  },
4599
  {
4600
  "name": "sdxl",
@@ -4886,7 +4888,8 @@
4886
  "input_price_per_1m": 0.4,
4887
  "output_price_per_1m": 2,
4888
  "currency": "USD",
4889
- "api_endpoint": "devstral-medium-latest"
 
4890
  },
4891
  {
4892
  "name": "Devstral Small 2",
@@ -4910,7 +4913,8 @@
4910
  "api_endpoint": "codestral-latest",
4911
  "capabilities": [
4912
  "tools"
4913
- ]
 
4914
  },
4915
  {
4916
  "name": "Mistral Small 3.2",
@@ -9568,7 +9572,9 @@
9568
  "input_price_per_1m": 0,
9569
  "output_price_per_1m": 0,
9570
  "currency": "USD",
9571
- "category": "FLUX.1"
 
 
9572
  },
9573
  {
9574
  "name": "flux-2-pro",
@@ -9582,7 +9588,7 @@
9582
  "output_price_per_1m": 0,
9583
  "currency": "USD",
9584
  "category": "FLUX.2",
9585
- "size_b": 12,
9586
  "hf_id": "black-forest-labs/FLUX.2-pro"
9587
  },
9588
  {
@@ -9623,7 +9629,7 @@
9623
  "output_price_per_1m": 0,
9624
  "currency": "USD",
9625
  "category": "FLUX.2",
9626
- "size_b": 12,
9627
  "hf_id": "black-forest-labs/FLUX.2-flex"
9628
  },
9629
  {
@@ -9677,7 +9683,7 @@
9677
  "output_price_per_1m": 0,
9678
  "currency": "USD",
9679
  "category": "FLUX.2",
9680
- "size_b": 12,
9681
  "hf_id": "black-forest-labs/FLUX.2-max"
9682
  },
9683
  {
 
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",
 
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",
 
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
  "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
  "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,
 
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",
 
1239
  "hf_private": true
1240
  },
1241
  {
1242
+ "name": "google/gemini-2.5-flash-lite",
1243
+ "type": "vision",
1244
  "input_price_per_1m": 0.1,
1245
+ "output_price_per_1m": 0.4,
1246
  "currency": "USD",
1247
+ "capabilities": [
1248
+ "vision",
1249
+ "reasoning"
1250
+ ],
1251
+ "context_window": 1048576,
1252
+ "hf_private": true
1253
  },
1254
  {
1255
+ "name": "google/gemini-2.0-flash-001",
1256
+ "type": "vision",
1257
  "input_price_per_1m": 0.1,
1258
+ "output_price_per_1m": 0.4,
1259
  "currency": "USD",
 
 
1260
  "capabilities": [
1261
+ "vision"
 
1262
  ],
1263
+ "context_window": 1048576,
1264
+ "hf_private": true
1265
  },
1266
  {
1267
  "name": "mistral/devstral-small-latest",
 
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",
 
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
  {
 
1436
  "hf_id": "mistralai/Mistral-Small-4-119B-2603"
1437
  },
1438
  {
1439
+ "name": "groq/openai/gpt-oss-120b",
1440
  "type": "chat",
1441
  "input_price_per_1m": 0.15,
1442
+ "output_price_per_1m": 0.75,
1443
  "currency": "USD",
1444
+ "context_window": 131072,
1445
+ "size_b": 120,
1446
  "capabilities": [
1447
+ "tools",
1448
  "reasoning"
1449
  ],
 
 
1450
  "hf_id": "openai/gpt-oss-120b"
1451
  },
1452
  {
 
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,
1529
  "currency": "USD",
1530
  "capabilities": [
1531
+ "vision",
1532
+ "reasoning"
1533
  ],
1534
+ "context_window": 400000,
1535
  "hf_private": true
1536
  },
1537
  {
 
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,
1583
  "currency": "USD",
 
1584
  "capabilities": [
1585
+ "vision"
 
1586
  ],
1587
+ "context_window": 2000000,
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
  {
1617
+ "name": "xai/grok-4-fast-non-reasoning",
1618
+ "type": "vision",
1619
+ "input_price_per_1m": 0.2,
1620
+ "output_price_per_1m": 0.5,
1621
  "currency": "USD",
1622
+ "context_window": 2000000,
1623
+ "capabilities": [
1624
+ "vision",
1625
+ "tools",
1626
+ "reasoning"
1627
+ ],
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",
 
1665
  ],
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",
 
1687
  ],
1688
  "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
1689
  },
 
 
 
 
 
 
 
 
 
 
 
 
1690
  {
1691
  "name": "azure/gpt-5-mini",
1692
  "type": "vision",
 
1724
  "context_window": 1048576,
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",
 
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
  "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
  "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,
 
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,
1910
  "currency": "USD",
1911
+ "capabilities": [
1912
+ "vision",
1913
+ "reasoning"
1914
+ ],
1915
+ "context_window": 1048576,
1916
+ "hf_private": true
1917
  },
1918
  {
1919
+ "name": "nebius/minimax/minimax-m2.1",
1920
  "type": "chat",
1921
  "input_price_per_1m": 0.3,
1922
+ "output_price_per_1m": 1.2,
1923
  "currency": "USD",
1924
+ "capabilities": [
1925
+ "reasoning"
1926
+ ],
1927
+ "context_window": 200000,
1928
+ "hf_id": "MiniMaxAI/MiniMax-M2.1",
1929
+ "size_b": 228.7
1930
  },
1931
  {
1932
  "name": "nebius/deepseek-ai/DeepSeek-V3.2",
 
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,
1949
  "currency": "USD",
1950
  "capabilities": [
1951
+ "vision",
1952
  "reasoning"
1953
  ],
1954
  "context_window": 200000,
1955
+ "hf_id": "MiniMaxAI/MiniMax-M2.7",
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",
 
2064
  "context_window": 1047576,
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"
2077
+ ],
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",
2094
  "type": "vision",
 
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",
 
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": [
2182
  "tools"
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",
2198
  "input_price_per_1m": 0.5,
2199
  "output_price_per_1m": 2,
 
2206
  "hf_private": true
2207
  },
2208
  {
2209
+ "name": "vertex/gemini-3-flash-preview",
2210
  "type": "vision",
2211
  "input_price_per_1m": 0.5,
2212
  "output_price_per_1m": 3,
 
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",
2265
  "input_price_per_1m": 0.5,
2266
+ "output_price_per_1m": 2,
2267
  "currency": "USD",
 
2268
  "capabilities": [
2269
+ "vision",
2270
+ "reasoning"
2271
  ],
2272
+ "context_window": 131072,
2273
+ "hf_private": true
2274
  },
2275
  {
2276
+ "name": "google/gemini-3-flash-preview",
2277
+ "type": "vision",
2278
  "input_price_per_1m": 0.5,
2279
+ "output_price_per_1m": 3,
2280
  "currency": "USD",
 
2281
  "capabilities": [
2282
+ "vision",
2283
  "reasoning"
2284
  ],
2285
+ "context_window": 1048576,
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",
 
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",
2424
+ "type": "vision",
2425
+ "input_price_per_1m": 0.6,
2426
+ "output_price_per_1m": 2.4,
2427
  "currency": "USD",
 
2428
  "capabilities": [
2429
+ "vision",
2430
+ "reasoning"
2431
  ],
2432
+ "context_window": 200000,
2433
+ "hf_id": "MiniMaxAI/MiniMax-M2.5",
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",
 
2475
  "size_b": 1026.5
2476
  },
2477
  {
2478
+ "name": "moonshot/kimi-k2-thinking",
2479
  "type": "chat",
2480
  "input_price_per_1m": 0.6,
2481
+ "output_price_per_1m": 2.5,
2482
  "currency": "USD",
2483
+ "context_window": 131072,
2484
  "capabilities": [
2485
+ "tools"
 
2486
  ],
2487
+ "hf_id": "moonshotai/Kimi-K2-Instruct",
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
  "hf_id": "zai-org/GLM-4.5",
2554
  "size_b": 358.3
2555
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2556
  {
2557
  "name": "novita/microsoft/wizardlm-2-8x22b",
2558
  "type": "chat",
 
2636
  "context_window": 400000,
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",
2650
  "type": "chat",
 
2685
  "hf_id": "deepseek-ai/DeepSeek-R1-0528",
2686
  "size_b": 684.5
2687
  },
 
 
 
 
 
 
 
 
 
2688
  {
2689
  "name": "deepinfra/deepseek-ai/DeepSeek-R1",
2690
  "type": "chat",
 
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,
 
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,
 
2791
  "capabilities": [
2792
  "vision"
2793
  ],
2794
+ "context_window": 1048576,
2795
  "hf_private": true
2796
  },
2797
  {
 
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,
2852
+ "currency": "USD",
2853
+ "capabilities": [
2854
+ "vision"
2855
+ ],
2856
+ "context_window": 200000,
2857
+ "hf_private": true
2858
+ },
2859
  {
2860
  "name": "zai/GLM-5",
2861
  "type": "vision",
 
2896
  "hf_id": "moonshotai/Kimi-K2-Instruct",
2897
  "size_b": 1026.5
2898
  },
 
 
 
 
 
 
 
 
 
 
 
 
2899
  {
2900
  "name": "azure/o4-mini",
2901
  "type": "vision",
 
2944
  "context_window": 200000,
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",
2958
  "type": "chat",
 
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
  {
 
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",
 
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,
 
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,
 
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,
 
3164
  ],
3165
  "hf_private": true
3166
  },
 
 
 
 
 
 
 
 
 
 
 
 
3167
  {
3168
  "name": "azure/openai-responses/gpt-5.2-codex",
3169
  "type": "vision",
 
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
  {
3206
+ "name": "openai-responses/gpt-5.3-chat",
3207
  "type": "vision",
3208
  "input_price_per_1m": 1.75,
3209
  "output_price_per_1m": 14,
 
3212
  "vision",
3213
  "reasoning"
3214
  ],
3215
+ "context_window": 128000,
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,
 
3225
  "vision",
3226
  "reasoning"
3227
  ],
3228
+ "context_window": 400000,
3229
  "hf_private": true
3230
  },
3231
  {
 
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
  "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,
 
3264
  "vision",
3265
  "reasoning"
3266
  ],
3267
+ "context_window": 400000,
3268
  "hf_private": true
3269
  },
3270
  {
 
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,
 
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,
 
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,
3350
  "currency": "USD",
3351
  "capabilities": [
3352
  "vision",
3353
  "reasoning"
3354
  ],
3355
+ "context_window": 1048576,
 
 
 
 
 
 
 
 
 
3356
  "hf_private": true
3357
  },
3358
  {
 
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,
 
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,
3389
  "currency": "USD",
3390
  "capabilities": [
3391
+ "vision"
 
3392
  ],
3393
+ "context_window": 1047576,
3394
  "hf_private": true
3395
  },
3396
  {
 
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
 
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",
 
3499
  "hf_private": true
3500
  },
3501
  {
3502
+ "name": "openai/gpt-4o-2024-08-06",
3503
  "type": "vision",
3504
  "input_price_per_1m": 2.5,
3505
  "output_price_per_1m": 10,
 
3510
  "context_window": 128000,
3511
  "hf_private": true
3512
  },
3513
+ {
3514
+ "name": "openai/gpt-5.4",
3515
+ "type": "vision",
3516
+ "input_price_per_1m": 2.5,
3517
+ "output_price_per_1m": 15,
3518
+ "currency": "USD",
3519
+ "capabilities": [
3520
+ "vision",
3521
+ "reasoning"
3522
+ ],
3523
+ "context_window": 1050000,
3524
+ "hf_private": true
3525
+ },
3526
  {
3527
  "name": "openai/gpt-4o-2024-05-13",
3528
  "type": "vision",
 
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,
 
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,
 
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,
3567
  "currency": "USD",
3568
  "capabilities": [
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
  "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,
 
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
  "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,
 
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,
3670
  "currency": "USD",
 
3671
  "capabilities": [
3672
+ "vision",
3673
+ "reasoning"
3674
  ],
3675
+ "context_window": 200000,
3676
  "hf_private": true
3677
  },
3678
  {
 
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,
 
3701
  "hf_private": true
3702
  },
3703
  {
3704
+ "name": "bedrock/claude-3-7-sonnet",
3705
  "type": "vision",
3706
  "input_price_per_1m": 3,
3707
  "output_price_per_1m": 15,
 
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,
3721
  "currency": "USD",
3722
  "capabilities": [
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
  "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,
 
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
  "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
  "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,
 
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,
 
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
  {
3846
+ "name": "bedrock/claude-opus-4-5",
3847
  "type": "vision",
3848
  "input_price_per_1m": 5,
3849
  "output_price_per_1m": 25,
 
3856
  "hf_private": true
3857
  },
3858
  {
3859
+ "name": "bedrock/claude-opus-4-6",
3860
  "type": "vision",
3861
  "input_price_per_1m": 5,
3862
+ "output_price_per_1m": 25,
3863
  "currency": "USD",
3864
  "capabilities": [
3865
+ "vision",
3866
+ "reasoning"
3867
  ],
3868
+ "context_window": 1000000,
3869
  "hf_private": true
3870
  },
3871
  {
 
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,
 
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,
3928
  "currency": "USD",
3929
  "capabilities": [
3930
  "vision",
3931
  "reasoning"
3932
  ],
3933
+ "context_window": 200000,
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
  "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",
 
3960
  "hf_private": true
3961
  },
3962
  {
3963
+ "name": "coding/claude-opus-4-20250514",
3964
  "type": "vision",
3965
  "input_price_per_1m": 15,
3966
+ "output_price_per_1m": 75,
3967
  "currency": "USD",
3968
  "capabilities": [
3969
  "vision",
 
3973
  "hf_private": true
3974
  },
3975
  {
3976
+ "name": "bedrock/claude-opus-4",
3977
  "type": "vision",
3978
  "input_price_per_1m": 15,
3979
  "output_price_per_1m": 75,
 
4594
  "currency": "USD",
4595
  "capabilities": [
4596
  "image-gen"
4597
+ ],
4598
+ "size_b": 12,
4599
+ "hf_id": "black-forest-labs/FLUX.1-dev"
4600
  },
4601
  {
4602
  "name": "sdxl",
 
4888
  "input_price_per_1m": 0.4,
4889
  "output_price_per_1m": 2,
4890
  "currency": "USD",
4891
+ "api_endpoint": "devstral-medium-latest",
4892
+ "ollama_id": "devstral-2"
4893
  },
4894
  {
4895
  "name": "Devstral Small 2",
 
4913
  "api_endpoint": "codestral-latest",
4914
  "capabilities": [
4915
  "tools"
4916
+ ],
4917
+ "ollama_id": "codestral"
4918
  },
4919
  {
4920
  "name": "Mistral Small 3.2",
 
9572
  "input_price_per_1m": 0,
9573
  "output_price_per_1m": 0,
9574
  "currency": "USD",
9575
+ "category": "FLUX.1",
9576
+ "size_b": 12,
9577
+ "hf_id": "black-forest-labs/FLUX.1-dev"
9578
  },
9579
  {
9580
  "name": "flux-2-pro",
 
9588
  "output_price_per_1m": 0,
9589
  "currency": "USD",
9590
  "category": "FLUX.2",
9591
+ "size_b": 32,
9592
  "hf_id": "black-forest-labs/FLUX.2-pro"
9593
  },
9594
  {
 
9629
  "output_price_per_1m": 0,
9630
  "currency": "USD",
9631
  "category": "FLUX.2",
9632
+ "size_b": 32,
9633
  "hf_id": "black-forest-labs/FLUX.2-flex"
9634
  },
9635
  {
 
9683
  "output_price_per_1m": 0,
9684
  "currency": "USD",
9685
  "category": "FLUX.2",
9686
+ "size_b": 32,
9687
  "hf_id": "black-forest-labs/FLUX.2-max"
9688
  },
9689
  {
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, 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 => {
@@ -66,6 +66,7 @@ function updateProviderModels(providers, providerName, models) {
66
  // But preserve these if newModel doesn't have them
67
  size_b: newModel.size_b || existing.size_b,
68
  hf_id: newModel.hf_id || existing.hf_id,
 
69
  hf_private: newModel.hf_private ?? existing.hf_private,
70
  capabilities: (newModel.capabilities && newModel.capabilities.length > 0)
71
  ? newModel.capabilities
@@ -154,6 +155,26 @@ function findOrMatch(modelName, orIndex) {
154
  return null;
155
  }
156
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
157
  // Fetch total_parameters from Hugging Face Hub API (Metadata)
158
  async function fetchHFSize(hfId) {
159
  if (!hfId || hfId.includes(' ') || !hfId.includes('/')) return { error: 'Invalid ID' };
@@ -165,11 +186,18 @@ async function fetchHFSize(hfId) {
165
 
166
  // Check various common metadata locations for total parameters
167
  let params = data.safetensors?.total || data.config?.total_parameters || data.config?.model_type_params;
 
 
168
  if (!params && data.cardData?.model_details?.parameters) {
169
  const match = data.cardData.model_details.parameters.match(/([\d.]+)\s*[Bb]/);
170
  if (match) params = parseFloat(match[1]) * 1_000_000_000;
171
  }
172
 
 
 
 
 
 
173
  if (!params) return { error: 'No parameter data in Hub metadata' };
174
 
175
  const b = params / 1_000_000_000;
@@ -183,6 +211,34 @@ async function fetchHFSize(hfId) {
183
  }
184
  }
185
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
186
  const EMBEDDER_KEYWORDS = ['embed', 'bge', 'gte', 'e5', 'stella', 'minilm', 'multilingual-mpnet'];
187
 
188
  // Link common models to their HF IDs when naming is non-standard
@@ -207,6 +263,7 @@ const MANUAL_HF_ID_MAP = {
207
  'qwen vl max': 'Qwen/Qwen2-VL-72B-Instruct',
208
  // FLUX detailed mappings
209
  'flux 1 dev': 'black-forest-labs/FLUX.1-dev',
 
210
  'flux 2 dev': 'black-forest-labs/FLUX.2-dev',
211
  'flux 2 klein 4b': 'black-forest-labs/FLUX.2-klein-4B',
212
  'flux 2 klein 9b': 'black-forest-labs/FLUX.2-klein-9B',
@@ -219,12 +276,6 @@ const MANUAL_HF_ID_MAP = {
219
  'flux 1 1 pro ultra': 'black-forest-labs/FLUX.1-pro',
220
  'flux 1 fill pro': 'black-forest-labs/FLUX.1-pro',
221
  'flux 1 kontext max': 'black-forest-labs/FLUX.1-pro',
222
- // Mistral mappings
223
- 'mistral large 2407': 'mistralai/Mistral-Large-Instruct-2407',
224
- 'mistral large latest': 'mistralai/Mistral-Large-Instruct-2407',
225
- 'mistral large 2': 'mistralai/Mistral-Large-Instruct-2407',
226
- 'mistral large 2411': 'mistralai/Mistral-Large-Instruct-2411',
227
- 'mistral large 3': 'mistralai/Mistral-Large-Instruct-2411',
228
  };
229
 
230
  const MANUAL_SIZE_MAP = {
@@ -272,9 +323,11 @@ async function propagateExtraData(data) {
272
  let propagatedSize = 0;
273
  let autoTagged = 0;
274
  let hfSizeFetched = 0;
 
275
 
276
  // We'll collect models missing size that have a clear HF-id-like name
277
  const hfLookupQueue = [];
 
278
 
279
  for (const provider of data.providers) {
280
  for (const model of provider.models || []) {
@@ -286,11 +339,13 @@ async function propagateExtraData(data) {
286
  }
287
 
288
  // 1. STRUCTURED LOOKUP: Match size by hf_id if available (Benchmark gold-standard)
289
- if (!model.size_b && model.hf_id) {
 
290
  if (MANUAL_SIZE_MAP[model.hf_id]) {
291
  model.size_b = MANUAL_SIZE_MAP[model.hf_id];
292
  propagatedSize++;
293
- } else {
 
294
  const size = hfIdToSize.get(model.hf_id.toLowerCase());
295
  if (size) { model.size_b = size; propagatedSize++; }
296
  }
@@ -345,10 +400,13 @@ async function propagateExtraData(data) {
345
  else if (n.includes('gemma 2 2b') || n.includes('gemma2 2b')) { model.size_b = 2; propagatedSize++; }
346
  }
347
 
348
- // 6. QUEUE: Still missing size? Try Hub API metadata lookup
349
- // Skip models that we've previously marked as private/unauthorized
350
- if (!model.size_b && !model.hf_private && (model.name.includes('/') || model.hf_id)) {
351
- hfLookupQueue.push(model);
 
 
 
352
  }
353
  }
354
  }
@@ -399,9 +457,37 @@ async function propagateExtraData(data) {
399
  console.log(` βœ“ Total ${hfSizeFetched} new sizes from HF metadata`);
400
  }
401
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
402
  if (autoTagged > 0) console.log(`Auto-tagged ${autoTagged} image-gen/embedding models.`);
403
  if (propagatedCaps > 0) console.log(`Propagated capabilities to ${propagatedCaps} models.`);
404
- if (propagatedSize + hfSizeFetched > 0) console.log(`Enriched size data for ${propagatedSize + hfSizeFetched} models.`);
405
  }
406
 
407
  async function runFetcher(fetcher, data) {
 
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 => {
 
66
  // But preserve these if newModel doesn't have them
67
  size_b: newModel.size_b || existing.size_b,
68
  hf_id: newModel.hf_id || existing.hf_id,
69
+ ollama_id: newModel.ollama_id || existing.ollama_id,
70
  hf_private: newModel.hf_private ?? existing.hf_private,
71
  capabilities: (newModel.capabilities && newModel.capabilities.length > 0)
72
  ? newModel.capabilities
 
155
  return null;
156
  }
157
 
158
+ // Estimate parameters from config.json (vLLM style fallback)
159
+ function estimateParams(config) {
160
+ if (!config) return null;
161
+ const h = config.hidden_size || config.d_model || config.n_embd;
162
+ const l = config.num_hidden_layers || config.n_layer;
163
+ const v = config.vocab_size;
164
+ const i = config.intermediate_size || config.d_ff;
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;
174
+ }
175
+ return null;
176
+ }
177
+
178
  // Fetch total_parameters from Hugging Face Hub API (Metadata)
179
  async function fetchHFSize(hfId) {
180
  if (!hfId || hfId.includes(' ') || !hfId.includes('/')) return { error: 'Invalid ID' };
 
186
 
187
  // Check various common metadata locations for total parameters
188
  let params = data.safetensors?.total || data.config?.total_parameters || data.config?.model_type_params;
189
+
190
+ // Fallback: cardData
191
  if (!params && data.cardData?.model_details?.parameters) {
192
  const match = data.cardData.model_details.parameters.match(/([\d.]+)\s*[Bb]/);
193
  if (match) params = parseFloat(match[1]) * 1_000_000_000;
194
  }
195
 
196
+ // Fallback: vLLM-style estimation from config
197
+ if (!params && data.config) {
198
+ params = estimateParams(data.config);
199
+ }
200
+
201
  if (!params) return { error: 'No parameter data in Hub metadata' };
202
 
203
  const b = params / 1_000_000_000;
 
211
  }
212
  }
213
 
214
+ // Fetch parameter info from Ollama Registry
215
+ async function fetchOllamaMetadata(modelName) {
216
+ const slug = modelName.toLowerCase().replace(/[^a-z0-9]/g, '-').replace(/-+/g, '-');
217
+ const url = `https://registry.ollama.ai/v2/library/${slug}/manifests/latest`;
218
+ try {
219
+ const data = await getJson(url, {
220
+ headers: { Accept: 'application/vnd.docker.distribution.manifest.v2+json' },
221
+ retries: 1
222
+ });
223
+ if (!data.config?.digest) return null;
224
+
225
+ // Fetch the config blob
226
+ const configUrl = `https://registry.ollama.ai/v2/library/${slug}/blobs/${data.config.digest}`;
227
+ const config = await getJson(configUrl, { retries: 1 });
228
+
229
+ const info = config.model_info || {};
230
+ const count = info['general.parameter_count'] || info['parameter_count'];
231
+ if (count) {
232
+ const b = count / 1_000_000_000;
233
+ const size = b < 1 ? Math.round(b * 100) / 100 : Math.round(b * 10) / 10;
234
+ return { size, ollama_id: slug };
235
+ }
236
+ return { ollama_id: slug }; // Found model but no size
237
+ } catch (e) {
238
+ return null;
239
+ }
240
+ }
241
+
242
  const EMBEDDER_KEYWORDS = ['embed', 'bge', 'gte', 'e5', 'stella', 'minilm', 'multilingual-mpnet'];
243
 
244
  // Link common models to their HF IDs when naming is non-standard
 
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',
267
  'flux 2 dev': 'black-forest-labs/FLUX.2-dev',
268
  'flux 2 klein 4b': 'black-forest-labs/FLUX.2-klein-4B',
269
  'flux 2 klein 9b': 'black-forest-labs/FLUX.2-klein-9B',
 
276
  'flux 1 1 pro ultra': 'black-forest-labs/FLUX.1-pro',
277
  'flux 1 fill pro': 'black-forest-labs/FLUX.1-pro',
278
  'flux 1 kontext max': 'black-forest-labs/FLUX.1-pro',
 
 
 
 
 
 
279
  };
280
 
281
  const MANUAL_SIZE_MAP = {
 
323
  let propagatedSize = 0;
324
  let autoTagged = 0;
325
  let hfSizeFetched = 0;
326
+ let ollamaFetched = 0;
327
 
328
  // We'll collect models missing size that have a clear HF-id-like name
329
  const hfLookupQueue = [];
330
+ const ollamaLookupQueue = [];
331
 
332
  for (const provider of data.providers) {
333
  for (const model of provider.models || []) {
 
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
  }
 
400
  else if (n.includes('gemma 2 2b') || n.includes('gemma2 2b')) { model.size_b = 2; propagatedSize++; }
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
  }
 
457
  console.log(` βœ“ Total ${hfSizeFetched} new sizes from HF metadata`);
458
  }
459
 
460
+ // 8. OLLAMA REGISTRY: Inspect parameter info (Final fallback for common models)
461
+ const uniqueOllama = [...new Set(ollamaLookupQueue.map(m => m.name.split('/').pop().replace(/:free$/, '')))].slice(0, 50);
462
+ if (uniqueOllama.length > 0) {
463
+ console.log(`\n Ollama: inspecting registry for ${uniqueOllama.length} models...`);
464
+ const nameToOllama = new Map();
465
+ for (let i = 0; i < uniqueOllama.length; i++) {
466
+ const name = uniqueOllama[i];
467
+ process.stdout.write(` [${i+1}/${uniqueOllama.length}] ${name.padEnd(50)} `);
468
+ const res = await fetchOllamaMetadata(name);
469
+ if (res) {
470
+ nameToOllama.set(name, res);
471
+ process.stdout.write(res.size ? `βœ“ ${res.size}B\n` : `βœ“ (exists)\n`);
472
+ } else {
473
+ process.stdout.write(`βœ—\n`);
474
+ }
475
+ await new Promise(r => setTimeout(r, 50));
476
+ }
477
+ for (const model of ollamaLookupQueue) {
478
+ const name = model.name.split('/').pop().replace(/:free$/, '');
479
+ const res = nameToOllama.get(name);
480
+ if (res) {
481
+ if (res.size && !model.size_b) { model.size_b = res.size; ollamaFetched++; }
482
+ if (res.ollama_id) model.ollama_id = res.ollama_id;
483
+ }
484
+ }
485
+ console.log(` βœ“ Total ${ollamaFetched} new sizes from Ollama`);
486
+ }
487
+
488
  if (autoTagged > 0) console.log(`Auto-tagged ${autoTagged} image-gen/embedding models.`);
489
  if (propagatedCaps > 0) console.log(`Propagated capabilities to ${propagatedCaps} models.`);
490
+ if (propagatedSize + hfSizeFetched + ollamaFetched > 0) console.log(`Enriched size data for ${propagatedSize + hfSizeFetched + ollamaFetched} models.`);
491
  }
492
 
493
  async function runFetcher(fetcher, data) {
src/App.css CHANGED
@@ -86,12 +86,12 @@ header p {
86
  background-color: var(--white);
87
  cursor: pointer;
88
  }
89
-
90
  .table-container {
91
  background-color: var(--white);
92
  border-radius: 0.75rem;
93
  box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.1), 0 2px 4px -1px rgba(0, 0, 0, 0.06);
94
- overflow: hidden;
 
95
  border: 1px solid var(--border-color);
96
  }
97
 
@@ -99,8 +99,10 @@ table {
99
  width: 100%;
100
  border-collapse: collapse;
101
  text-align: left;
 
102
  }
103
 
 
104
  th {
105
  background-color: #f1f5f9;
106
  padding: 1rem;
 
86
  background-color: var(--white);
87
  cursor: pointer;
88
  }
 
89
  .table-container {
90
  background-color: var(--white);
91
  border-radius: 0.75rem;
92
  box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.1), 0 2px 4px -1px rgba(0, 0, 0, 0.06);
93
+ overflow-x: auto;
94
+ margin-bottom: 2rem;
95
  border: 1px solid var(--border-color);
96
  }
97
 
 
99
  width: 100%;
100
  border-collapse: collapse;
101
  text-align: left;
102
+ min-width: 800px; /* Ensure table doesn't squash too much */
103
  }
104
 
105
+
106
  th {
107
  background-color: #f1f5f9;
108
  padding: 1rem;
src/App.tsx CHANGED
@@ -17,6 +17,7 @@ interface Model {
17
  capabilities?: string[]
18
  display_name?: string
19
  hf_id?: string
 
20
  }
21
 
22
  interface Provider {
@@ -76,10 +77,18 @@ interface BenchmarkEntry {
76
  aider_pass_rate?: number; // 0-1, first-pass success on 133 coding tasks
77
  // Artificial Analysis (artificialanalysis.ai)
78
  aa_id?: string;
 
 
79
  aa_intelligence?: number; // 0-100 intelligence index
 
 
80
  aa_mmlu_pro?: number;
81
  aa_gpqa?: number;
82
  aa_livecodebench?: number;
 
 
 
 
83
  aa_tokens_per_s?: number;
84
  aa_latency_s?: number;
85
  }
@@ -154,6 +163,9 @@ function App() {
154
  if (b.lb_name) nameMap.set(normalizeName(b.lb_name), b);
155
  // Chatbot Arena display name
156
  if (b.arena_name) nameMap.set(normalizeName(b.arena_name), b);
 
 
 
157
  }
158
  return { nameMap, hfIdMap };
159
  }, [liveBenchmarks]);
@@ -535,6 +547,14 @@ function App() {
535
  </a>
536
  </div>
537
  )}
 
 
 
 
 
 
 
 
538
  {model.capabilities && model.capabilities.length > 0 && (
539
  <div className="tooltip-row"><strong>Caps:</strong> {model.capabilities.join(', ')}</div>
540
  )}
 
17
  capabilities?: string[]
18
  display_name?: string
19
  hf_id?: string
20
+ ollama_id?: string
21
  }
22
 
23
  interface Provider {
 
77
  aider_pass_rate?: number; // 0-1, first-pass success on 133 coding tasks
78
  // Artificial Analysis (artificialanalysis.ai)
79
  aa_id?: string;
80
+ aa_name?: string;
81
+ aa_slug?: string;
82
  aa_intelligence?: number; // 0-100 intelligence index
83
+ aa_coding?: number; // 0-100 coding index
84
+ aa_math?: number; // 0-100 math index
85
  aa_mmlu_pro?: number;
86
  aa_gpqa?: number;
87
  aa_livecodebench?: number;
88
+ aa_hle?: number;
89
+ aa_scicode?: number;
90
+ aa_math_500?: number;
91
+ aa_aime?: number;
92
  aa_tokens_per_s?: number;
93
  aa_latency_s?: number;
94
  }
 
163
  if (b.lb_name) nameMap.set(normalizeName(b.lb_name), b);
164
  // Chatbot Arena display name
165
  if (b.arena_name) nameMap.set(normalizeName(b.arena_name), b);
166
+ // Artificial Analysis name and slug
167
+ if (b.aa_name) nameMap.set(normalizeName(b.aa_name), b);
168
+ if (b.aa_slug) nameMap.set(normalizeName(b.aa_slug), b);
169
  }
170
  return { nameMap, hfIdMap };
171
  }, [liveBenchmarks]);
 
547
  </a>
548
  </div>
549
  )}
550
+ {model.ollama_id && (
551
+ <div className="tooltip-row">
552
+ <strong>Ollama:</strong>
553
+ <a href={`https://ollama.com/library/${model.ollama_id}`} target="_blank" rel="noopener noreferrer" className="hf-link">
554
+ {model.ollama_id} β†—
555
+ </a>
556
+ </div>
557
+ )}
558
  {model.capabilities && model.capabilities.length > 0 && (
559
  <div className="tooltip-row"><strong>Caps:</strong> {model.capabilities.join(', ')}</div>
560
  )}