Mistral released the weights of an improved Mistral Large 4 on October 29, moving beyond the API preview announced earlier this month. The trillion-parameter model accepts text and images, supports tool calling, and offers low, medium and high reasoning settings. Developers can now download official FP8 and NVFP4 checkpoints.
What changed after the preview
The release brings more than downloadable files. Mistral reports that additional training lifted Terminal-Bench 4.0 performance by 13 percentage points to 41%, while reducing output tokens by 30%. Its API now costs $0.70 per million input tokens and $2.10 per million output tokens.
These benchmark gains are Mistral’s reported results, not a guarantee for every repository. Lower token usage in a benchmark also does not establish the total cost of a self-hosted deployment.
Why it matters for builders
For developers building coding and defensive-security agents, open weights create another option for keeping model execution alongside private code, customizing behavior and avoiding dependence on one hosted API.
There are practical limits: this is a substantial infrastructure deployment, not a lightweight laptop model. The Mistral Ecosystem Open-Weight License also requires a separate commercial license for businesses operating a Model-as-a-Service offering. Downloadable weights do not mean unrestricted use.
For SunMarc App Labs, the useful next step would be a bounded bug-fix evaluation: compare patch quality, test results and total cost against the current coding workflow before considering adoption.
Our earlier coverage of Mistral Large 4’s API preview described the promise of a later weights release. This announcement makes that deployment option available. The evaluation above is a proposal, not an announced SunMarc integration.