AgentsModel picker
Local models you can run on your own hardware, for coding, images, video and voice. These are our editorial notes with facts taken from each model's own page. For rankings, we link to independent leaderboards.
17 models. Showing every tier. Model names link to the official Hugging Face page. New to downloading models? Read Hugging Face for beginners.
| Model | Runs with | Smallest tier | License | Our notes |
|---|---|---|---|---|
| Qwen2.5-Coder 7B Instruct | Ollama, LM Studio | Laptop | Apache-2.0 | 7.6B parameters. A 4-bit (Q4) GGUF is about 5GB, so it runs on a 16GB laptop. Ollama tag: qwen2.5-coder:7b. Good for autocomplete and small edits. |
| Qwen3 8B | Ollama, LM Studio | Laptop | Apache-2.0 | 8B general model with a "thinking" mode that helps on code reasoning. Q4 is about 5GB. Ollama tag: qwen3:8b. |
| Phi-4-mini Instruct | Ollama, LM Studio | Laptop | MIT | 3.8B parameters, about 2.5GB at Q4. The pick for 8GB machines or when you want very fast replies; weaker on large codebases. |
| DeepSeek-Coder-V2-Lite Instruct | Ollama, LM Studio | Laptop | DeepSeek model license (commercial use allowed) | 16B mixture-of-experts with only about 2.4B active per token, so it is quick. Q4 is about 10GB: wants a 32GB laptop or Mac. |
| gpt-oss-20b | Ollama, LM Studio | Laptop | Apache-2.0 | OpenAI open-weight model, 21B total with about 3.6B active. Its card says it runs within 16GB of memory; plan on a 32GB laptop/Mac for comfort. Strong at tool use. |
| Gemma 3 12B | Ollama, LM Studio | Laptop | Gemma terms of use | 12B general model from Google. Q4 is about 8GB; fits a 16–32GB Mac or an 8–12GB GPU. Accept the license on Hugging Face before downloading. |
| Qwen3-Coder 30B-A3B Instruct | Ollama, LM Studio | Gaming GPU | Apache-2.0 | 30.5B total, about 3.3B active, built for agentic coding tools. Q4 is about 18–19GB: fits a 24GB GPU or a 32GB+ Mac. Ollama tag: qwen3-coder:30b. |
| Qwen2.5-Coder 32B Instruct | Ollama, LM Studio | Gaming GPU | Apache-2.0 | Dense 32.8B coding model. Q4 is about 20GB, so it needs a 24GB card (RTX 3090/4090) or a 32GB+ Mac. |
| Devstral Small (2507) | Ollama, LM Studio | Gaming GPU | Apache-2.0 | 23.6B agentic coding model from Mistral, made for tools like OpenHands. Q4 is about 14GB; fits a 16–24GB GPU. |
| Qwen3 32B | Ollama, LM Studio | Gaming GPU | Apache-2.0 | Dense 32.8B general model, strong at reasoning about code. Q4 is about 20GB, so plan on a 24GB GPU. |
| Gemma 3 27B | Ollama, LM Studio | Gaming GPU | Gemma terms of use | 27B general model; Q4 is about 17GB, fits a 24GB GPU. Gated: accept the terms on Hugging Face first. |
| Codestral 22B | Ollama, LM Studio | Gaming GPU | Mistral Non-Production License | 22B code model with fill-in-the-middle autocomplete. Q4 is about 13GB. Not licensed for commercial use: personal and research only. |
| gpt-oss-120b | Ollama, LM Studio | Workstation | Apache-2.0 | 117B total, about 5.1B active. Its card says it fits on a single 80GB GPU; also runs on a 96–128GB Mac. Fast for its size. |
| GLM-4.5-Air | LM Studio, llama.cpp | Workstation | MIT | 110B mixture-of-experts (about 12B active) aimed at agents and coding. Q4 is roughly 65–70GB: a 96–128GB Mac or 2–3 large GPUs. |
| Llama 3.3 70B Instruct | Ollama, LM Studio | Workstation | Llama 3.3 Community License | Dense 70B general model. Q4 is about 42GB: two 24GB GPUs, one 48GB card, or a 64GB+ Mac. Gated: accept Meta’s license first. |
| Qwen2.5 72B Instruct | Ollama, LM Studio | Workstation | Qwen license (check the card for commercial terms) | Dense 72.7B model. Q4 is about 47GB, so it wants 48GB+ of VRAM or a 64GB+ Mac. |
| Qwen3 235B-A22B Instruct (2507) | llama.cpp, LM Studio | Workstation | Apache-2.0 | 235B total, 22B active. Even a 3-bit GGUF is over 100GB, so this needs a 128GB+ Mac or a multi-GPU rig. The top end of what runs at home. |
Coding agents that can drive these local models: Aider, Cline, OpenCode, Kilo Code and Qwen Code. Prefer a cloud agent? Compare GitHub Copilot, Claude Code, OpenAI Codex, Gemini CLI, Cursor and Devin Desktop (ex-Windsurf) in local vs cloud coding.
Compare benchmark rankings (cloud and open models) on Artificial Analysis: language model leaderboards. We link to their leaderboards and don't republish their data.
Hardware tiers
How we sort models. A model listed for a tier runs there at a usable speed with a 4-bit (Q4) or similar quantized file, which is the size most people download. A rough rule of thumb, not a guarantee. Check each model page for its stated requirements. LM Studio recommends at least 16GB of RAM and 4GB of dedicated VRAM.
| Tier | Memory | What usually fits |
|---|---|---|
| Laptop | 16–32 GB RAM or unified memory (any Apple Silicon Mac), or a laptop/desktop GPU with 8 GB VRAM or less | Coding/chat models up to about 8B (or ~20B mixture-of-experts at 32 GB), SD 1.5/SDXL images, light video, all small TTS |
| Gaming GPU | A desktop GPU with 12–24 GB VRAM (e.g. RTX 3060 12GB up to RTX 3090/4090/5090), or a 32–64 GB Mac | Coding models up to about 32B at Q4, FLUX-class images, 5B-class video, expressive TTS |
| Workstation | 48 GB+ VRAM (one big card or several GPUs), or a Mac with 96–128 GB+ unified memory | 70B–235B models, 20B image models at full precision, 14B-class video at 720p |
Getting started
Want the model to do tasks, not just chat? See Local AI agents: OpenClaw, Hermes Desktop and VS Code, step by step. How to find, size and download any model: Hugging Face for beginners. New to local models? Start with Run AI models on your own computer: Ollama, LM Studio and Open WebUI.
For language and coding models, install Ollama or LM Studio and pull a model tag that fits your memory. For images and video, ComfyUI is the common open-source workspace. Always check a model's license before business use.