Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableMistral AI Blog→ original

Mistral Releases Weights for Large 4: 19.7% of Agent Patches Would Be Merged by Maintainers

Mistral has released the weights for Large 4 (Le Chonk), an open 1-trillion-parameter model for coding. On Cognition's closed FrontierCode 2.0 test, which asks whether a repository maintainer would accept an agent's patch, Large 4 scores 19.7%, while DeepSeek V4.1 Flash scores 19.3%. On the Coding Agent Index, it reaches 48%, matching the open-model leader GLM-5.3, but uses 47% fewer tokens per task. For a coding agent running on your own servers, that could mean a lower token bill.

AI-processed from Mistral AI Blog; edited by Hamidun News
Mistral Releases Weights for Large 4: 19.7% of Agent Patches Would Be Merged by Maintainers
Source: Mistral AI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceMistral AI Blog ↗
Editor-in-Chief
Terminal-Bench 4.041%, 13 percentage points above the preview, with 30% fewer output tokens
Cyber Index51%; refuses malicious cyber requests in 96.6% of cases
Three reasoning modesLow, Medium, and High (the default)

Mistral has published the weights for Mistral Large 4, which the company calls Le Chonk. It is a flagship multimodal model with 1 trillion parameters, trained from scratch on European infrastructure. You can download it, deploy it yourself, and fine-tune it without going through Mistral's API. The headline figure is 19.7% on Cognition's closed FrontierCode 2.0 test. The test is grounded in practice: an agent edits a real repository, and the evaluators check whether the code as a whole is good and whether a live project maintainer would have merged it. So nearly one in five of Large 4's patches would have made it into the codebase. Among open models, this is the best result in Mistral's comparison table.

All figures below come from Mistral itself. Some of the tests were run by third-party teams, but the article contains no independent re-verification.

Coding: Ahead of Its Peers and More Efficient with Tokens

A preview of the model came out earlier this month. The team then spent three more weeks on reinforcement learning (RL), in which the model receives a reward for correctly solved tasks. The results improved, and Mistral sees no sign of saturation.

On FrontierCode 2.0, Large 4 sits alongside DeepSeek V4.1 Flash (19.3%), GLM-5.3 (17.8%), and Kimi K3 (17.3%). The margin is small, but it is noticeable.

On the Artificial Analysis Coding Agent Index, Large 4 scores 48%. That is almost the same as GLM-5.3, the leader among open models. However, the model uses on average 47% fewer tokens per task. If an agent solves hundreds of tasks a day, the difference shows up in the inference bill.

In terminal scenarios, where the agent runs commands on its own, the model scored 41% on Terminal-Bench 4.0. That is 13 percentage points higher than the preview and the best result among open models in the comparison. At the same time, output tokens fell by 30%.

In Surge Code's blind pairwise comparisons, evaluators chose Large 4's answers more often than those of GLM-5.3 (64%), Kimi K3 (64%), and DeepSeek V4.1 Flash (56%). It lost only to Claude Opus 5.5, which won 78% of comparisons.

There are now three reasoning modes:

  • Low for simple tasks, replacing the previous no-reasoning mode;
  • Medium for everyday work;
  • High for the most complex tasks that require precision. This is the default mode.

Cybersecurity: Helping Defenders, Not Attackers

On the Artificial Analysis Cyber Index, the model scored 51%. The index checks three steps: finding a vulnerability in code, reproducing it and confirming it, and then fixing it without breaking anything. Mistral's chart includes models with higher scores, up to 57%. Mistral's emphasis lies elsewhere: on legitimate defensive tasks, Large 4 does not refuse without good reason.

Built-in restrictions separate defense from offense. The model refuses malicious cyber requests in 96.6% of cases on average. That is the highest rate among the open models Mistral has measured. The evaluations used requests from JailbreakBench, StrongREJECT, and AgentHarm. For comparison, in the same chart DeepSeek-V4-Pro scores about 51%, Gemma-4-31B about 84%, GLM-5.3 about 87%, and Kimi K3 about 94%.

Partners also show interest. Booz Allen writes that the model matched other leading cybersecurity solutions in the completeness of vulnerability detection at a more favorable cost. The agency ENISA is testing it on vulnerability discovery, code auditing, and analysis of defensive logs. Elastic already supports Mistral's open models in Elastic Security. Cisco, Cloudflare, CrowdStrike, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik, and Zscaler have also offered feedback. Almost everywhere, the same argument recurs: defenders need a model that they control themselves.

Sovereignty and Limitations

For banks, government agencies, and regulated industries in Europe, it matters that the model was trained entirely in Europe and that its provenance can be traced from start to release. Data does not need to be sent to a foreign jurisdiction, and the architecture can be studied and audited. Mistral calls this "sovereign AI."

There are caveats as well. The company itself published the test results, and FrontierCode 2.0 is closed, so it cannot be reproduced by hand. The leadership in coding is among open models only: Claude Opus 5.5 was stronger in the blind comparison. A model with 1 trillion parameters will require a serious GPU fleet.

The practical takeaway is simple. If you are building a coding or security agent and want to keep it on your own infrastructure, add Large 4 to your comparison. Run it on your own tasks and compare how many tokens it uses in Medium and High modes. That efficiency is likely to be the main argument for choosing it.

Story

Key Questions Answered

What are the key details on Coding: Ahead of Its Peers and More Efficient with Tokens?

A preview of the model came out earlier this month. The team then spent three more weeks on reinforcement learning (RL), in which the model receives a reward for correctly solved tasks. The results improved, and Mistral sees no sign of saturation. On FrontierCode 2.0, Large 4 sits alongside DeepSeek V4.1 Flash (19.3%), GLM-5.3 (17.8%), and Kimi K3 (17.3%). The margin is small, but it is noticeable. On the Artificial Analysis Coding Agent Index, Large 4 scores 48%.

What are the key details on Cybersecurity: Helping Defenders, Not Attackers?

On the Artificial Analysis Cyber Index, the model scored 51%. The index checks three steps: finding a vulnerability in code, reproducing it and confirming it, and then fixing it without breaking anything. Mistral's chart includes models with higher scores, up to 57%. Mistral's emphasis lies elsewhere: on legitimate defensive tasks, Large 4 does not refuse without good reason. Built-in restrictions separate defense from offense. The model refuses malicious cyber requests in 96.6% of cases on average.

What are the key details on Sovereignty and Limitations?

For banks, government agencies, and regulated industries in Europe, it matters that the model was trained entirely in Europe and that its provenance can be traced from start to release. Data does not need to be sent to a foreign jurisdiction, and the architecture can be studied and audited. Mistral calls this "sovereign AI." There are caveats as well. The company itself published the test results, and FrontierCode 2.0 is closed, so it cannot be reproduced by hand.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…