Simon Willison has released llm-mistral 0.16, the latest update to his plugin that connects the llm command-line tool to Mistral's API. The release, out on 6 October, adds support for reasoning models, including the newly announced Mistral Large 4, controlled with a new -o reasoning_effort option.
Under the hood, the plugin has switched to using Mistral's own official Python library rather than hand-rolled API calls, which should make it easier to keep pace with new Mistral features as they ship. There is one breaking change worth flagging for existing users: the safe_mode option has been removed. Anyone who relied on it needs to switch to -o safe_prompt 1 instead, which does the same job under Mistral's current API.
The release also extends Voxtral, Mistral's audio model family, so it now accepts local MP3 files directly rather than requiring a publicly accessible URL. That is a small change with an outsized effect on usability: anyone transcribing or querying private audio no longer needs to upload a file somewhere else first just to get a URL for it.
As with Willison's other provider plugins, the pattern is tight tracking of a fast-moving API. Mistral ships a new reasoning model and a revised safety parameter, and within days the command-line tool that wraps it catches up, keeping llm mistral usable as a daily driver rather than something that lags a release behind.
Source: Simon Willison
