Browse · 40 tracked
Local LLMs
We track 40 open-weight models from 7 developers — 7 MoE among them — with measured GGUF sizes per quant, each graded against all 50 machines we cover. “Runs on” counts machines where the model loads cleanly (grade B or better) at a quality quant with llama.cpp at 8K context — see how the math works.
| Model | Params | Architecture | Max context | Quants | Runs on |
|---|---|---|---|---|---|
| DeepSeek-R1 | 671B | MoE · 37B active | 160K | 4 | none |
| DeepSeek-V4-Flash | 284B | MoE · 13B active | 1024K | 4 | 2/50 |
| gpt-oss-120b | 117B | MoE · 5.1B active | 128K | 1 | 6/50 |
| Qwen2.5-72B-Instruct | 72.7B | dense | 32K | 4 | 7/50 |
| Llama-3.1-70B-Instruct | 70.6B | dense | 128K | 4 | 7/50 |
| Llama-3.3-70B-Instruct | 70.6B | dense | 128K | 4 | 7/50 |
| Qwen3.6-35B-A3B | 35B | MoE · 3B active | 256K | 4 | 16/50 |
| Qwen2.5-32B-Instruct | 32.8B | dense | 32K | 4 | 16/50 |
| QwQ-32B | 32.8B | dense | 40K | 4 | 16/50 |
| Qwen3-32B | 32.8B | dense | 40K | 4 | 16/50 |
| DeepSeek-R1-Distill-Qwen-32B | 32.8B | dense | 128K | 4 | 16/50 |
| Qwen3-30B-A3B | 30.5B | MoE · 3.3B active | 40K | 4 | 20/50 |
| Qwen3.6-27B | 27.8B | dense | 256K | 4 | 20/50 |
| gemma-3-27b-it | 27.4B | dense | 128K | 4 | 16/50 |
| gemma-2-27b-it | 27.2B | dense | 8K | 4 | 20/50 |
| gemma-4-26B-A4B-it | 26.5B | MoE · 3.8B active | 256K | 4 | 20/50 |
| Mistral-Small-24B-Instruct-2501 | 23.6B | dense | 32K | 4 | 21/50 |
| gpt-oss-20b | 21B | MoE · 3.6B active | 128K | 1 | 36/50 |
| DeepSeek-R1-Distill-Qwen-14B | 14.8B | dense | 128K | 4 | 36/50 |
| Qwen2.5-14B-Instruct | 14.8B | dense | 32K | 4 | 36/50 |
| Qwen3-14B | 14.8B | dense | 40K | 4 | 36/50 |
| Phi-4 | 14.7B | dense | 16K | 4 | 36/50 |
| Mistral-Nemo-Instruct-2407 | 12.3B | dense | 128K | 4 | 45/50 |
| gemma-3-12b-it | 12.2B | dense | 128K | 4 | 36/50 |
| Qwen3.5-9B | 9.7B | dense | 256K | 4 | 46/50 |
| gemma-2-9b-it | 9.2B | dense | 8K | 4 | 46/50 |
| Qwen3-8B | 8.2B | dense | 40K | 4 | 50/50 |
| DeepSeek-R1-Distill-Llama-8B | 8B | dense | 128K | 4 | 50/50 |
| Llama-3.1-8B-Instruct | 8B | dense | 128K | 4 | 50/50 |
| DeepSeek-R1-Distill-Qwen-7B | 7.6B | dense | 128K | 4 | 50/50 |
| Qwen2.5-7B-Instruct | 7.6B | dense | 32K | 4 | 50/50 |
| Mistral-7B-Instruct-v0.3 | 7.3B | dense | 32K | 4 | 50/50 |
| gemma-3-4b-it | 4.3B | dense | 128K | 4 | 50/50 |
| Phi-3.5-mini-instruct | 3.8B | dense | 128K | 4 | 50/50 |
| Llama-3.2-3B-Instruct | 3.2B | dense | 128K | 4 | 50/50 |
| Qwen2.5-3B-Instruct | 3.1B | dense | 32K | 4 | 50/50 |
| gemma-2-2b-it | 2.6B | dense | 8K | 4 | 50/50 |
| Qwen2.5-1.5B-Instruct | 1.5B | dense | 32K | 4 | 50/50 |
| Llama-3.2-1B-Instruct | 1.2B | dense | 128K | 4 | 50/50 |
| Qwen2.5-0.5B-Instruct | 0.5B | dense | 32K | 4 | 50/50 |
Params, layers, and context come from each repo's config.json; quant sizes are measured from the published GGUF files. “none” is an honest answer, not a data gap — some models simply don't fit consumer hardware.