Which model to take
The assistant works with many models, but not all of them build well. This page says which ones the app suggests, what a cloud model costs and how much memory a model on your own computer needs. The same advice is in Settings ▸ Assistant under Which model to take and under the model list of your own server.
What every model needs
- It must call tools. The assistant reads your catalog and changes the graph through tools. The model list marks each model tools or no tool calls. A model that cannot call tools is asked again without them: it then answers only from the graph it was sent, and the chat says "Answered without tools". It cannot build or offer drafts.
- A large enough context window: 16 384 tokens or more for a model on your own computer. See The context window.
- Bigger builds better. A small model answers simple questions but often loses its way in a long build.
Cloud models
For each cloud provider the app suggests two models: one To start (cheaper) and one For long builds (stronger). Prices are in US dollars per million tokens, as of 14 November 2026. The provider's own page has the current prices, and the provider's bill is the truth.
| Provider | To start | For long builds | Its web search |
|---|---|---|---|
| OpenAI | gpt-6-luna: $0.10 sent, $0.50 received | gpt-6.1-sol: $2 sent, $10 received | about $10 per 1000 searches |
| Anthropic | claude-haiku-5-5: $0.10 sent, $0.50 received | claude-sonnet-5-5: $2 sent, $10 received | about $10 per 1000 searches |
| Google Gemini | gemini-3.5-flash-lite: $0.30 sent, $2.50 received | gemini-4-hydrogen: $1.50 sent, $7.50 received | about $14 per 1000 searches |
What a request costs: "A short question is a fraction of a cent; a build of a dozen steps sends tens of thousands of tokens." The live model list of your key always wins: you can pick any model it offers.
An OpenAI-compatible service has its own prices; the app shows none for it. See OpenAI-compatible services.
A model on your own computer
"A model on your own computer costs nothing per question, and nothing leaves the computer." The app suggests these Ollama models, loaded with a window of 16 384 tokens:
| Model | Video memory (or, slower, system memory) |
|---|---|
qwen3:8b | about 8 GB (about 5 GB to download) |
qwen3:14b | about 12 GB |
qwen3:30b | about 24 GB |
To start: install Ollama (ollama.com/download), then run ollama pull qwen3:8b. See Ollama. Other servers (LM Studio, the llama.cpp server, vLLM) can run the same models under their own names.