Sync 2026-09-13: 141 models

README.md

@@ -25,10 +25,10 @@ configs:
         path: models.csv
 ---
 
-# ModelFit — Local LLM Hardware Compatibility Dataset
+# ModelFit: Local LLM Hardware Compatibility Dataset
 
-An open dataset of **which local AI models (Ollama) fit which hardware** — by
-parameter size, quantization, minimum RAM, and estimated memory load — across
+An open dataset of **which local AI models (Ollama) fit which hardware**, by
+parameter size, quantization, minimum RAM, and estimated memory load, across
 Apple Silicon Macs, iPhones, and NVIDIA GPUs.
 
 Maintained by **[ModelFit](https://modelfit.io/)**. Browse it as an interactive
@@ -37,7 +37,7 @@ machine-readable source is
 **[modelfit.io/api/dataset](https://modelfit.io/api/dataset/)**.
 
 - **141 models** across 24 families (106 with a registry-verified local build, 35 cloud-only APIs tracked for comparison)
-- **License:** CC BY 4.0 — reuse freely, including commercially, with attribution
+- **License:** CC BY 4.0: reuse freely, including commercially, with attribution
 - **Updated:** see the `updated` field in `models.json` (regenerated from the live endpoint)
 
 ```python
@@ -59,19 +59,19 @@ ds.filter(lambda r: r["runsLocally"] and r["minRamGb"] <= 16).to_pandas()
 
 ## Columns
 
-- `model` — display name (e.g. `Qwen3.5 9B Instruct`)
-- `family` — model family (Qwen, Llama, Gemma, DeepSeek, …)
-- `params` — parameter count, in billions (`null` when the vendor does not disclose it, e.g. closed API models)
-- `quantization` — e.g. `Q4_K_M`
-- `minRamGb` — minimum unified memory / VRAM to load it
-- `estimatedLoadGb` — approximate memory footprint at this quantization
-- `kvKbPerToken` — exact fp16 KV-cache cost in KB per token for hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next), computed from the published HF config (full-attention layers × kv_heads × head_dim × 2 × 2 bytes; only full-attention layers cache KV). `null` for standard GQA models, whose KV is estimated by size class. Example: Qwen3.6 35B-A3B is 20 KB/token, so a full 262k-token fp16 cache is ~5 GB
-- `runsLocally` — `true` when a registry-verified Ollama build fits at least one consumer RAM tier tracked here (up to 256GB)
-- `openWeights` — `true` when the weights are publicly downloadable. Can be `true` while `runsLocally` is `false`: open-weight giants like NVIDIA Nemotron 3 Ultra (550B, ~190GB at 2-bit) or Kimi K2 exceed every consumer tier
-- `ggufDiy` — `true` when the weights are open and a ~Q4 GGUF (0.6 GB per billion parameters) fits a 256GB-class machine via llama.cpp, but no Ollama build exists (e.g. DeepSeek V4 Flash 284B, Xiaomi MiMo-V2-Flash 309B) — runnable DIY, not scored for local fit
-- `runtimes` — apps the model runs in (`ollama`, `llama.cpp`, `lm-studio`; pipe-separated in the CSV, empty for cloud rows)
-- `bestFor` — primary workloads
-- `ollamaCommand` — exact `ollama run …` command (local models)
+- `model`: display name (e.g. `Qwen3.5 9B Instruct`)
+- `family`: model family (Qwen, Llama, Gemma, DeepSeek, …)
+- `params`: parameter count, in billions (`null` when the vendor does not disclose it, e.g. closed API models)
+- `quantization`: e.g. `Q4_K_M`
+- `minRamGb`: minimum unified memory / VRAM to load it
+- `estimatedLoadGb`: approximate memory footprint at this quantization
+- `kvKbPerToken`: exact fp16 KV-cache cost in KB per token for hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next), computed from the published HF config (full-attention layers × kv_heads × head_dim × 2 × 2 bytes; only full-attention layers cache KV). `null` for standard GQA models, whose KV is estimated by size class. Example: Qwen3.6 35B-A3B is 20 KB/token, so a full 262k-token fp16 cache is ~5 GB
+- `runsLocally`: `true` when a registry-verified Ollama build fits at least one consumer RAM tier tracked here (up to 256GB)
+- `openWeights`: `true` when the weights are publicly downloadable. Can be `true` while `runsLocally` is `false`: open-weight giants like NVIDIA Nemotron 3 Ultra (550B, ~190GB at 2-bit) or Kimi K2 exceed every consumer tier
+- `ggufDiy`: `true` when the weights are open and a ~Q4 GGUF (0.6 GB per billion parameters) fits a 256GB-class machine via llama.cpp, but no Ollama build exists (e.g. DeepSeek V4 Flash 284B, Xiaomi MiMo-V2-Flash 309B) — runnable DIY, not scored for local fit
+- `runtimes`: apps the model runs in (`ollama`, `llama.cpp`, `lm-studio`; pipe-separated in the CSV, empty for cloud rows)
+- `bestFor`: primary workloads
+- `ollamaCommand`: exact `ollama run …` command (local models)
 
 ## Methodology
 
@@ -106,13 +106,13 @@ You may share and adapt the data, including commercially, **with attribution**:
 ## Cite
 
 ```
-ModelFit — Local LLM Hardware Compatibility Dataset.
+ModelFit: Local LLM Hardware Compatibility Dataset.
 https://modelfit.io/data/
 ```
 
 ---
 
-Built and maintained by [ModelFit](https://modelfit.io/) — find the best local
+Built and maintained by [ModelFit](https://modelfit.io/). Find the best local
 AI model for your Mac, iPhone, or GPU. Mirrors:
 [GitHub](https://github.com/modelfit/modelfit-hardware-dataset) ·
 [Hugging Face](https://huggingface.co/datasets/modelfit/modelfit-hardware-dataset).