So what about self-hosted models?

[pelag0s]

The documentation mentions qwen2.5-coder:32b and from my personal experience that is currently one of the best coding models, also for coding assistant tasks.

Did anyone try it out with PR-Agent? Or did you find an even better model which can be self-hosted locally?

And if yes, what context size is needed for the review functionality to work correctly?

(I also found a page which says for self-hosted models, fine tuning is needed. But I also saw that most models tested there were quite old. Is this still needed?)

[IsmaelMartinez]

The documented answer has not moved much. The Ollama section still uses ollama/qwen2.5-coder:32b as its worked example, and the practical requirement is that you raise OLLAMA_CONTEXT_LENGTH well above Ollama's 2048 default and keep custom_model_max_tokens in step with it, since the review prompt plus a real diff will not fit otherwise; the docs suggest 128000 for that model. Ollama Cloud API keys are supported as of PR #2278. On quality, the project's position is unchanged: local open models are suitable for experimentation and /ask, while commercial models are recommended for production review. That note is dated January 2025 and is due a refresh, which is worth a separate docs issue rather than keeping this thread open. Closing as answered.