I attempted to use Devstral locally with LMStudio, version 24B, due to the limited memory available on my Mac. I managed to find the configuration file, modify the provided and model options, but unfortunately, I was unable to resolve the issue. Subsequently, I discovered a GitHub issue where users reported successfully running Devstral: https://github.com/mistralai/mistral-vibe/issues/9. Based on this information, I concluded that there might be a problem with my setup. After several attempts, I still have been able to find a solution.
What was the issue? The problem was the context size, which was set to 4096. The vibe message already contained 4555 tokens, but it wasn’t mentioned in the logs, so it took some time to identify the cause. You can change the context size here. Additionally, you can change the port (which isn’t the issue, as suggested in the issue – you can leave it as 1234).
Now, when the Claude Code limit will be reached, I can switch to local Devstral and continue working. At least, I hope I can (the token/s are not great either).