AI models
The right model for each task.
Nestor combines open-weight models installed on your side and, if you allow it, external models such as Claude. In automatic mode, it picks the cheapest model that can handle the task.
Open-weight models, on your side
For conversation and everyday tasks, Nestor relies first on models installed on your hardware.
- Served by Ollama or vLLM, chosen to fit your machine's memory (for example from the Qwen, Gemma, Granite, Mistral or gpt-oss families).
- For real use: a graphics card with at least 12 GB of video memory, or an Apple Silicon Mac with 16 GB of unified memory or more.
- A model installed on your side is not billed per token.
External models, if you allow them
Some tasks call for the most capable models. Your organisation decides whether to open them.
- Claude (Anthropic) is recommended to build agents.
- Other providers (OpenAI, Gemini) can be allowed.
- What is sent is flagged and logged, after automatic masking of some personal data.
Automatic routing
In automatic mode, Nestor picks the model for each request.
- The request is classified without calling a model.
- Among the models able to handle the task, the cheapest is chosen; at equal cost, the internal model comes first.
- If you pick a model yourself, that choice is never replaced.
Build with a model, run without
That is where the token savings come from.
- Building an agent, and each change to it, calls a model.
- Running an agent written by Nestor calls no model: 0 tokens.
- “Internal only” mode builds with a local model: slower, it may not manage complex tasks.