PLATE GPULocal models
Every model here runs on your machine, for free.
Osmium Coder talks to Ollama, LM Studio, and local GGUF files over localhost. No API key, no account for the model itself, and nothing metered — the compute is yours. These 16 are the curated picks with VRAM figures we have verified; any other Ollama tag works too.
- Curated models
- 16 Curated models
- Smallest
- 2 GB Smallest
- Largest
- 40 GB Largest
- API keys
- 0 API keys
Will it run on your GPU?
Pick your hardware and we will sort the catalog by what actually fits. Osmium does the same check on your real machine at install time, including system RAM and free disk.
7 of 16 models run on this. We would start you on qwen2.5-coder:7b.
- qwen2.5-coder:7b7B · 6 GB VRAM · 4.5 GB disk · 33K ctxFits
- qwen2.5-coder:3b3B · 3 GB VRAM · 2.3 GB disk · 33K ctxFits
- qwen2.5-coder:1.5b1.5B · 2 GB VRAM · 2 GB disk · 33K ctxFits
- llama3.1:8b8B · 6 GB VRAM · 4.5 GB disk · 128K ctxFits
- mistral:7b7B · 5 GB VRAM · 3.8 GB disk · 33K ctxFits
- qwen2.5vl:7b7B · 7 GB VRAM · 5.3 GB disk · 33K ctxTight
- llama3.2-vision:11b11B · 8 GB VRAM · 6 GB disk · 131K ctxTight
- qwen2.5-coder:14b14B · 10 GB VRAM · 7.5 GB disk · 33K ctxToo big
- qwen2.5-coder:32b32B · 20 GB VRAM · 15 GB disk · 33K ctxToo big
- deepseek-coder-v2:16b16B · 12 GB VRAM · 9 GB disk · 128K ctxToo big
- codestral:22b22B · 14 GB VRAM · 10.5 GB disk · 33K ctxToo big
- devstral:latest24B · 16 GB VRAM · 12 GB disk · 128K ctxToo big
- llama3.3:70b70B · 40 GB VRAM · 30 GB disk · 128K ctxToo big
- qwen2.5:14b14B · 10 GB VRAM · 7.5 GB disk · 33K ctxToo big
- gemma3:12b12B · 9 GB VRAM · 6.8 GB disk · 128K ctxToo big
- llava:13b13B · 10 GB VRAM · 7.5 GB disk · 4K ctxToo big
Checked against GPU memory only. The app also checks system RAM and free disk on your actual machine.
Coding
Coding models
Writes and edits code. Use these for Agent and Architect.
- qwen2.5-coder:1.5bQwen2.5 Coder 1.5B1.5B · 2 GB VRAM · 2 GB disk · 33K ctxFits 4 GB GPUs. Tools and fill-in-the-middle. A tiny local coder — not for big refactors.
- qwen2.5-coder:3bQwen2.5 Coder 3B3B · 3 GB VRAM · 2.3 GB disk · 33K ctxThe default local coder on 4-6 GB cards. Tools and fill-in-the-middle. Clearly stronger than 1.5B.
- qwen2.5-coder:7bQwen2.5 Coder 7B7B · 6 GB VRAM · 4.5 GB disk · 33K ctxThe best default local coder on 8-16 GB. Tools and fill-in-the-middle. This is the one most people should start with.
- qwen2.5-coder:14bQwen2.5 Coder 14B14B · 10 GB VRAM · 7.5 GB disk · 33K ctxThe sweet spot around 12 GB of VRAM. Handles bigger refactors than 7B.
- qwen2.5-coder:32bQwen2.5 Coder 32B32B · 20 GB VRAM · 15 GB disk · 33K ctxThe closest local stand-in for a cloud coder. Wants a 24 GB class GPU.
- deepseek-coder-v2:16bDeepSeek Coder V2 16B16B · 12 GB VRAM · 9 GB disk · 128K ctxLong context and code-first. Weaker at long prose explanations.
- codestral:22bCodestral 22B22B · 14 GB VRAM · 10.5 GB disk · 33K ctxMistral's coder. Solid fill-in-the-middle, which is what Tab uses.
- devstral:latestDevstral24B · 16 GB VRAM · 12 GB disk · 128K ctxAgent-oriented coder. Reach for it when Qwen 14B is too small and 32B will not fit.
General
General models
Explains and reasons. Use these for Ask and Analyze.
- qwen2.5:14bQwen2.5 14B14B · 10 GB VRAM · 7.5 GB disk · 33K ctxA better teacher than the coder SKU. Use it for architecture write-ups and Analyze mode.
- llama3.1:8bLlama 3.1 8B8B · 6 GB VRAM · 4.5 GB disk · 128K ctxFast general chat. Not a specialist coder — prefer Qwen Coder for edits.
- llama3.3:70bLlama 3.3 70B70B · 40 GB VRAM · 30 GB disk · 128K ctxThe best local explainer if you have 48 GB or more. Slow.
- mistral:7bMistral 7B7B · 5 GB VRAM · 3.8 GB disk · 33K ctxSmall general model. Fine for drafts, not for agent loops.
- gemma3:12bGemma 3 12B12B · 9 GB VRAM · 6.8 GB disk · 128K ctxStrong explanations and a long context for a 12B.
Vision
Vision models
Reads screenshots and diagrams you attach. None of them generate images.
- llama3.2-vision:11bLlama 3.2 Vision 11B11B · 8 GB VRAM · 6 GB disk · 131K ctxReads screenshots and diagrams you attach. It does not generate images.
- llava:13bLLaVA 13B13B · 10 GB VRAM · 7.5 GB disk · 4K ctxThe classic local vision model. Short context, and weak at text-only coding.
- qwen2.5vl:7bQwen2.5-VL 7B7B · 7 GB VRAM · 5.3 GB disk · 33K ctxThe best small local reader for UI screenshots. Text out only.
Scores are guidance, not benchmarks
The coding and explaining scores on these pages come from our own use while building Osmium. They are there to help you choose between two models that both fit your GPU, and they are not benchmark results. VRAM figures assume a 4-bit quantisation, which is what Ollama pulls by default.