# Local models with the Assistant You can use a model running on your computer to work on the open design in SolidCraft's desktop Assistant. See [Agent chat](agent-chat.md) for the panel's controls. ## What you need - Ollama, LM Studio, or another OpenAI-compatible server running on your computer. - A model that supports **tool calling**: this lets it ask SolidCraft to read or change the design. On Ollama's model page, look for the **tools** tag, or run `ollama show ` and look for `tools` under capabilities. Tool support alone does not guarantee good results; see the table below. - Enough graphics memory (VRAM) for the model and its context (the conversation and tool instructions). See the table below for sizes. ## Set it up 1. Start your server and load or download the model you want to use. 2. In SolidCraft, open **Preferences > AI > Add agent…** and choose the **OpenAI-compatible API**. Choose Ollama, LM Studio, or a custom URL. 3. Enter the server's base URL, such as `http://localhost:11434/v1` for Ollama, and the model name. Use **Test**, then add the endpoint. 4. Open the **Assistant** and select your endpoint. Choose the model in the model list if the server provides one. 5. Choose permissions in Preferences or the chat's permission menu: **Ask** asks before changes and file actions; **Edits** allows model edits but still asks before file actions and replacing the design; **All** allows actions without approval, including saving, overwriting, and replacing the design. ## Which models work? These results were measured with the standard test (v1) in [#4851](https://github.com/tangency-labs/solidcraft/issues/4851). The table in that issue is the up-to-date one. | Model | Size / quant | Passed (v1) | |---|---|---| | Muse Glimmer 30B | Unsloth UD-Q3_K_XL, 12.4 GB | **11/11** | | Qwen3.8 27B | Unsloth UD-Q4_K_XL | **11/11** | | gemma4 (latest, 7.5B) | Q4_K_M, 6.1 GB | **11/11** | | gemma4:12b | Q4_K_M | 10/11 | | gpt-oss:20b | MXFP4, 12.8 GB | 10/11 | | qwen3:14b | Q4_K_M, 8.6 GB | 8/11 | | qwen3:8b | Q4_K_M | 7/11 | | qwen2.5:7b | Q4_K_M, 4.4 GB | 4/11 | | qwen2.5-coder:7b | Q4_K_M, 4.4 GB | 0/11 (writes every call as text) | | gemma3:4b | Q4_K_M | 0/11 (no tool support: HTTP 400) | | deepseek-r1:14b | Q4_K_M | 0/11 (never calls a tool) | For about 8 GB of graphics memory, start with `gemma4` (the default tag, 6.1 GB). With 16 GB, Muse Glimmer 30B or gpt-oss:20b also work well. ## Troubleshooting - **The model replies with JSON or code, but nothing happens.** It is not calling tools well. Try another model from the results above. - **“HTTP 400 … does not support tools.”** The model lacks tool support. Choose one with the **tools** capability. - **Empty replies.** The context may be too small for the tool instructions. Increase the server's context size (`num_ctx` in Ollama), then try a new chat. - **Slow answers.** The model may be too big for your VRAM and spilling work to the CPU. Try a smaller model or a smaller quantization. ## Measure a model yourself Follow [the standard test in #4851](https://github.com/tangency-labs/solidcraft/issues/4851) to compare your model with these results. It gives the exact prompts, starting designs, pass rules, and a template for sharing results.