Best open models that run on a CPU
You can run smaller language models and vision-language models on a CPU using GGUF (via llama.cpp or Ollama), ONNX Runtime, or llamafile. Speed depends on your RAM bandwidth; models needing 1–8 GB run on most laptops with 16 GB of RAM.
Last updated 3 November 2026 · refreshed after every Hugging Face sync · 15 models
| # | Model | Size | Memory | License | Downloads (30 days) | Likes |
|---|---|---|---|---|---|---|
| 1 | Ornith-1.5-9B-GGUF ornith-ai | 9B · GGUF | ~6 GB | MIT | 5.4M | 625 |
| 2 | Ornith-1.0-9B-GGUF ornith-ai | 9B · GGUF | ~6 GB | MIT | 2.4M | 699 |
| 3 | SmolLM2-135M-Instruct HuggingFaceTB | 135M · ONNX | ~1 GB | Apache 2.0 | 827.1K | 468 |
| 4 | LFM2.5-2.6B-GGUF LiquidAI | 2.6B · GGUF | ~2 GB | Custom (other) | 782K | 386 |
| 5 | Qwythos-9B-Claude-Mythos-5-1M-GGUF empero-ai | 9B · GGUF | ~6 GB | Apache 2.0 | 234.9K | 2.9K |
| 6 | Sharp-Spark-X2.5-4B-GGUF peculiar-ragdoll | 4B · GGUF | ~3 GB | Apache 2.0 | 831.3K | 153 |
| 7 | gemma-7b | 7B · GGUF | ~5 GB | Gemma license | 136.7K | 3.5K |
| 8 | gemma-4-12B-coder-fable5-composer2.5-v1-GGUF yuxinlu1 | 12B · GGUF | ~8 GB | Apache 2.0 | 110.7K | 3K |
| 9 | Spark-X2.6-4B-GGUF XHToken | 4B · GGUF | ~3 GB | Apache 2.0 | 355.9K | 275 |
| 10 | JiRackUltra_1b CMSManhattan | 1B · GGUF | ~1 GB | MIT | 891.3K | 19 |
| 11 | NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF unsloth | 4B · GGUF | ~3 GB | Custom (other) | 281.5K | 157 |
| 12 | LFM3-1.7B-GGUF LiquidAI | 1.7B · GGUF | ~2 GB | Custom (other) | 174.2K | 199 |
| 13 | Mellum2.1-12B-A2.5B-Thinking-GGUF JetBrains | 12B · GGUF | ~8 GB | Apache 2.0 | 152K | 132 |
| 14 | NVIDIA-Nemotron-3.5-Bolt-E4B-4GB nvidia | 4B · GGUF | ~3 GB | Custom (other) | 92.7K | 344 |
| 15 | Ternary-Bonsai-2-9B-gguf prism-ml | 9B · GGUF | ~6 GB | Apache 2.0 | 111.6K | 124 |
How we rank
Language and vision-language models in GGUF, MLX, ONNX, or llamafile format whose published weights need about 1–8 GB. We include models seen on Hugging Face in the last 14 days with at least 10,000 downloads in the last 30 days. Score = log10(30-day downloads) + ½ × log10(likes + 1), so popularity leads and likes break near-ties. We keep one entry per model name and leave out adult and uncensored builds. Memory is estimated from parameter count and the published file format (16-bit ≈ 2 bytes per parameter, 4-bit ≈ 0.5) plus 20% headroom; long contexts need more. A quantised copy of a bigger model may also fit — check the model page.