Mistral Large 4 Open Weights: 1 Trillion Parameters Built for Coders and Defenders

On 29 October 2026 Mistral released the weights of Mistral Large 4 (Le Chonk): a 1-trillion-parameter, natively multimodal model trained from scratch on European infrastructure. Mistral calls it the best open-weight model for coding and reports 19.7% on Cognition's private FrontierCode 2.0

Tool Interface

Interactive tool will be available soon

Features

✓ A 1-trillion-parameter natively multimodal flagship whose weights can be downloaded and self-hosted, with Mistral stressing European training and traceable lineage
✓ Mistral says three more weeks of reinforcement learning after the preview moved coding and defensive cybersecurity substantially, with no sign of saturation
✓ Scores 19.7% on Cognition's private FrontierCode 2.0 benchmark, which judges end-to-end code quality and whether a repository maintainer would merge the change
✓ Better reasoning efficiency: the release note says developers get more performance at lower inference cost
✓ Ships several reasoning modes so you can trade cost against capability per task

How to Use

Read the weight and licence notes on the official release page first instead of assuming open-source defaults
Download the weights per the official model card and run a minimal inference pass to confirm your hardware and stack
Pick a reasoning mode: the cheap one for routine work, the heavy one for multi-file refactors
Validate against real issues in your own repository rather than leaderboard ranks alone

FAQ

When did it ship?

29 October 2026. Official release page: https://mistral.ai/news/mistral-large-4-oss . The model first appeared as a public preview earlier in the month and the improved weights landed three weeks later.

Is it really 1 trillion parameters?

The official page describes a 1-trillion-parameter natively multimodal model. Active parameters, VRAM requirements and other deployment details belong to the official model card, so do not size a GPU plan from the headline total.

Can I use it commercially?

Do not guess. The release page covers the weights and their conditions and the Hugging Face model card adds more. Read the licence line by line before choosing self-hosting or the hosted API.

What does 19.7% on FrontierCode 2.0 mean?

Mistral explains it as Cognition's private benchmark measuring end-to-end code quality and whether a repository maintainer would merge the agent's change. It is a score on that benchmark, not a claim that the model solves 19.7% of tasks in some absolute sense.

How is it different from the preview, then?

Mistral says the extra reinforcement learning sharply improved coding and defensive cybersecurity, cut inference cost through better reasoning efficiency, and that gains are still accelerating with no saturation in sight.