Sync 2026-10-21: 161 models

README.md

@@ -36,7 +36,7 @@ table at **[modelfit.io/data](https://modelfit.io/data/)**; the canonical
 machine-readable source is
 **[modelfit.io/api/dataset](https://modelfit.io/api/dataset/)**.
 
-- **156 models** across 27 families (116 with a registry-verified local build, 23 open-weight models without a consumer-tier Ollama build, 17 closed cloud APIs tracked for comparison)
+- **161 models** across 27 families (121 with a registry-verified local build, 23 open-weight models without a consumer-tier Ollama build, 17 closed cloud APIs tracked for comparison)
 - **License:** CC BY 4.0: reuse freely, including commercially, with attribution
 - **Updated:** see the `updated` field in `models.json` (regenerated from the live endpoint)
 

models.csv

@@ -111,7 +111,7 @@ MiniMax M3,MiniMax,428,API,0,0,,false,true,false,,"Coding, Agentic, Long context
 Qwen3.7-Plus,Qwen,,API,0,0,,false,false,false,,"Multimodal, Agentic, Vision",
 NVIDIA Nemotron 3 Ultra,Nemotron,550,API,0,0,,false,true,false,,"Reasoning, Agentic, Long context",
 Xiaomi MiMo-V2-Flash,MiMo,309,API,0,0,,false,true,true,,"Reasoning, Coding, Agentic tasks",
-NVIDIA Nemotron Cascade 2 30B-A3B,Nemotron,30,Q6_K,36,24,,true,true,false,ollama|llama.cpp|lm-studio,"Reasoning, Math, Agentic tasks",ollama run nemotron-cascade-2:30b
+NVIDIA Nemotron Cascade 2 30B-A3B,Nemotron,30,Q6_K,36,24,6,true,true,false,ollama|llama.cpp|lm-studio,"Reasoning, Math, Agentic tasks",ollama run nemotron-cascade-2:30b
 Poolside Laguna XS.2,Laguna,33,Q4_K_M,36,23,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, long-horizon software engineering",ollama run laguna-xs.2:q4_K_M
 Cohere North Mini Code,North,30,Q4_K_M,32,19,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, repo-scale software engineering",ollama run north-mini-code-1.0:q4_K_M
 GPT-OSS 120B,GPT-OSS,117,MXFP4,96,65.4,,true,true,false,ollama|llama.cpp|lm-studio,"Reasoning, Coding, Agents",ollama run gpt-oss:120b
@@ -133,20 +133,25 @@ Laguna S 2.1,Laguna,118,Q4_K_M,128,96,,true,true,false,ollama|llama.cpp|lm-studi
 Ornith 1.0 9B,Ornith,9,Q4_K_M,8,5.6,,true,true,false,ollama|llama.cpp|lm-studio,Agentic coding on small machines,ollama run ornith:9b
 Ornith 1.0 35B,Ornith,35,Q4_K_M,32,21.2,,true,true,false,ollama|llama.cpp|lm-studio,Agentic coding,ollama run ornith:35b
 Qwen3.8-Flash-Next,Qwen,125,Q4_K_XL,192,111.3,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Reasoning, Multimodal",
+Qwen3.8-Flash-Next (IQ4_XS),Qwen,125,IQ4_XS,128,93.7,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Reasoning, Multimodal",
+Qwen3.8-Flash-Next (Q3),Qwen,125,Q3_K_XL,128,90,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Reasoning, Multimodal",
 Qwen3.8 2.4T A95B,Qwen,2400,API,0,0,,false,true,false,,"Frontier reasoning, Agentic coding",
 Granite 4.2 3B,Granite,3.7,Q4_K_M,4,2.1,,true,true,false,ollama|llama.cpp|lm-studio,"Chat, Coding",ollama run granite4.2:3b
 Granite 4.2 8B,Granite,8.8,Q4_K_M,8,5,,true,true,false,ollama|llama.cpp|lm-studio,"Chat, Coding",ollama run granite4.2:8b
 Granite 4.2 30B,Granite,29.3,Q4_K_M,24,16.5,,true,true,false,ollama|llama.cpp|lm-studio,"Coding, Chat",ollama run granite4.2:30b
-Nemotron 3.5 Lightning 30B-A3B,Nemotron,30,Q4_K_M,36,23.7,6,true,true,false,ollama|llama.cpp|lm-studio,"Agentic, Coding, Long context",ollama run nemotron-3.5-lightning:30b
+Nemotron 3.5 Lightning 30B-A3B,Nemotron,30,Q4_K_M,36,25.4,6,true,true,false,ollama|llama.cpp|lm-studio,"Agentic, Coding, Long context",ollama run nemotron-3.5-lightning:30b
+Nemotron 3.5 Lightning 30B-A3B (IQ4_XS),Nemotron,30,IQ4_XS,32,18.9,6,true,true,false,ollama|llama.cpp|lm-studio,"Agentic, Coding, Long context",
 Muse Glimmer 30B,Muse,29.8,Q4_K_M,32,16.9,,true,true,false,ollama|llama.cpp|lm-studio,Chat,ollama run muse-glimmer:30b
 MiniCPM-V 4.5 8B,MiniCPM,8.7,Q4_K_M,12,5.7,,true,true,false,ollama|llama.cpp|lm-studio,"Vision, Multimodal",ollama run minicpm-v4.5:8b
 GLM-5.3,Zhipu,753,API,0,0,,false,true,false,,"Coding, Reasoning, Agentic",
 GLM-5.3 Flash,Zhipu,321,API,0,0,,false,true,true,,"Chat, Coding",
 GLM-4.7-Flash 30B-A3B,Zhipu,30,Q4_K_M,32,19,,true,true,false,ollama|llama.cpp|lm-studio,"Coding, Agentic tasks, Reasoning",ollama run glm-4.7-flash
-Nemotron 3 Super 120B-A12B,Nemotron,120,Q4_K_M,128,86.8,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic workflows, Long context, Reasoning",ollama run nemotron-3-super:120b
+Nemotron 3 Super 120B-A12B,Nemotron,120,Q4_K_M,128,86.8,8,true,true,false,ollama|llama.cpp|lm-studio,"Agentic workflows, Long context, Reasoning",ollama run nemotron-3-super:120b
+Nemotron 3 Super 120B-A12B (IQ4_XS),Nemotron,120,IQ4_XS,96,64.5,8,true,true,false,ollama|llama.cpp|lm-studio,"Agentic workflows, Long context, Reasoning",
 Olmo 3.1 32B Instruct,Olmo,32,Q4_K_M,32,19.5,,true,true,false,ollama|llama.cpp|lm-studio,"General chat, Research, Fully open",ollama run olmo-3.1:32b
 MiMo-V2.6-Distill-Qwen-9B,MiMo,9,Q4_K_M,16,5.8,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Small machines",
 Ling-3.0-flash,Ling,124,Q4_K_M,128,77.8,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Long context, Reasoning",
+Ling-3.0-flash (IQ4_XS),Ling,124,IQ4_XS,96,68.7,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Long context, Reasoning",
 Xiaomi MiMo-V2.6-Flash,MiMo,309,API,0,0,,false,true,true,,"Agentic coding, Multimodal, Long context",
 Xiaomi MiMo-V2.6-Pro,MiMo,1020,API,0,0,,false,true,false,,"Frontier agentic coding, Multimodal",
 Devstral Small 2 24B,Mistral,24,Q4_K_M,24,15.2,,true,true,false,ollama|llama.cpp|lm-studio,"Agentic coding, Software engineering",ollama run devstral-small-2:24b

models.json

@@ -11,10 +11,10 @@
     "note": "A model fits a device when its estimatedLoadGb is within the memory budget: ~70% of unified memory up to 32GB, scaling linearly to ~85% at 128GB and above (high-RAM Macs can wire more memory to the GPU). Q4_K_M needs roughly 0.6 GB per billion parameters.",
     "fieldSemantics": "runsLocally = a registry-verified Ollama build fits at least one consumer RAM tier tracked here (up to 256GB). openWeights = the weights are publicly downloadable. Some open-weight giants (e.g. NVIDIA Nemotron 3 Ultra, 550B, ~190GB at 2-bit) are openWeights: true but runsLocally: false because no consumer tier fits them."
   },
-  "updated": "2026-10-13",
+  "updated": "2026-10-21",
   "counts": {
-    "models": 156,
-    "local": 116,
+    "models": 161,
+    "local": 121,
     "cloud": 40,
     "families": 27
   },
@@ -2022,7 +2022,7 @@
       "quantization": "Q6_K",
       "minRamGb": 36,
       "estimatedLoadGb": 24,
-      "kvKbPerToken": null,
+      "kvKbPerToken": 6,
       "runsLocally": true,
       "openWeights": true,
       "ggufDiy": false,
@@ -2429,6 +2429,44 @@
       "bestFor": "Agentic coding, Reasoning, Multimodal",
       "ollamaCommand": null
     },
+    {
+      "model": "Qwen3.8-Flash-Next (IQ4_XS)",
+      "family": "Qwen",
+      "params": 125,
+      "quantization": "IQ4_XS",
+      "minRamGb": 128,
+      "estimatedLoadGb": 93.7,
+      "kvKbPerToken": null,
+      "runsLocally": true,
+      "openWeights": true,
+      "ggufDiy": false,
+      "runtimes": [
+        "ollama",
+        "llama.cpp",
+        "lm-studio"
+      ],
+      "bestFor": "Agentic coding, Reasoning, Multimodal",
+      "ollamaCommand": null
+    },
+    {
+      "model": "Qwen3.8-Flash-Next (Q3)",
+      "family": "Qwen",
+      "params": 125,
+      "quantization": "Q3_K_XL",
+      "minRamGb": 128,
+      "estimatedLoadGb": 90,
+      "kvKbPerToken": null,
+      "runsLocally": true,
+      "openWeights": true,
+      "ggufDiy": false,
+      "runtimes": [
+        "ollama",
+        "llama.cpp",
+        "lm-studio"
+      ],
+      "bestFor": "Agentic coding, Reasoning, Multimodal",
+      "ollamaCommand": null
+    },
     {
       "model": "Qwen3.8 2.4T A95B",
       "family": "Qwen",
@@ -2507,7 +2545,7 @@
       "params": 30,
       "quantization": "Q4_K_M",
       "minRamGb": 36,
-      "estimatedLoadGb": 23.7,
+      "estimatedLoadGb": 25.4,
       "kvKbPerToken": 6,
       "runsLocally": true,
       "openWeights": true,
@@ -2520,6 +2558,25 @@
       "bestFor": "Agentic, Coding, Long context",
       "ollamaCommand": "ollama run nemotron-3.5-lightning:30b"
     },
+    {
+      "model": "Nemotron 3.5 Lightning 30B-A3B (IQ4_XS)",
+      "family": "Nemotron",
+      "params": 30,
+      "quantization": "IQ4_XS",
+      "minRamGb": 32,
+      "estimatedLoadGb": 18.9,
+      "kvKbPerToken": 6,
+      "runsLocally": true,
+      "openWeights": true,
+      "ggufDiy": false,
+      "runtimes": [
+        "ollama",
+        "llama.cpp",
+        "lm-studio"
+      ],
+      "bestFor": "Agentic, Coding, Long context",
+      "ollamaCommand": null
+    },
     {
       "model": "Muse Glimmer 30B",
       "family": "Muse",
@@ -2614,7 +2671,7 @@
       "quantization": "Q4_K_M",
       "minRamGb": 128,
       "estimatedLoadGb": 86.8,
-      "kvKbPerToken": null,
+      "kvKbPerToken": 8,
       "runsLocally": true,
       "openWeights": true,
       "ggufDiy": false,
@@ -2626,6 +2683,25 @@
       "bestFor": "Agentic workflows, Long context, Reasoning",
       "ollamaCommand": "ollama run nemotron-3-super:120b"
     },
+    {
+      "model": "Nemotron 3 Super 120B-A12B (IQ4_XS)",
+      "family": "Nemotron",
+      "params": 120,
+      "quantization": "IQ4_XS",
+      "minRamGb": 96,
+      "estimatedLoadGb": 64.5,
+      "kvKbPerToken": 8,
+      "runsLocally": true,
+      "openWeights": true,
+      "ggufDiy": false,
+      "runtimes": [
+        "ollama",
+        "llama.cpp",
+        "lm-studio"
+      ],
+      "bestFor": "Agentic workflows, Long context, Reasoning",
+      "ollamaCommand": null
+    },
     {
       "model": "Olmo 3.1 32B Instruct",
       "family": "Olmo",
@@ -2683,6 +2759,25 @@
       "bestFor": "Agentic coding, Long context, Reasoning",
       "ollamaCommand": null
     },
+    {
+      "model": "Ling-3.0-flash (IQ4_XS)",
+      "family": "Ling",
+      "params": 124,
+      "quantization": "IQ4_XS",
+      "minRamGb": 96,
+      "estimatedLoadGb": 68.7,
+      "kvKbPerToken": null,
+      "runsLocally": true,
+      "openWeights": true,
+      "ggufDiy": false,
+      "runtimes": [
+        "ollama",
+        "llama.cpp",
+        "lm-studio"
+      ],
+      "bestFor": "Agentic coding, Long context, Reasoning",
+      "ollamaCommand": null
+    },
     {
       "model": "Xiaomi MiMo-V2.6-Flash",
       "family": "MiMo",