Mistral AI has unveiled its Large 4 model weights, now available to the public, showcasing significant advancements in AI capabilities.
The French AI firm Mistral has made the weights of its flagship model, Large 4 (Le Chonk), publicly available, now totaling over 1 trillion parameters.
Available on Hugging Face, the model previously operated only as a public preview via API.
Utilizing a Mixture of Experts architecture, Large 4 processes each token with approximately 49 billion parameters—up to 52 billion with additional layers—allowing for reduced computational resource usage without sacrificing size.
This model can handle context sizes of up to 1 million tokens, supports text and image inputs, and comprehends more than 160 languages. It features three reasoning modes for varying complexity: Low, Medium, and High.
Following its initial preview, Mistral continued to enhance the model for three weeks using reinforcement learning, leading to improved performance in programming, cybersecurity, and agency tasks.
In recent tests, Large 4 showed a 13 percentage point improvement in programming tasks via the Terminal-Bench 4.0, reaching 41% accuracy while cutting token output by 30%.
According to the Artificial Analysis Coding Agent Index, the model scored 48%, closely trailing the Chinese model GLM-5.3. Mistral noted that Large 4 used 47% fewer tokens on similar tasks.
In the private benchmark FrontierCode 2.0, which assesses code quality and the likelihood of acceptance by project maintainers, Large 4 achieved a score of 19.7%.
In a blind comparison, it was preferred less frequently than competitors like GLM-5.3 and Claude Opus 5.5, though it fared better than Kimi K3 and DeepSeek V4.1 Flash.
Mistral has emphasized cybersecurity, with Large 4 scoring 51% on the Artificial Analysis Cyber Index, which measures its ability to identify and fix code vulnerabilities. The model rejected 96.6% of malicious cybersecurity-related queries, and its resistance to attacks increased from 93.3% to 97.1% in the B3 Agent Security Benchmark.
Some evaluation results were generated in-house and others by external firms, with a full technical report expected to be released later.
While Mistral acknowledges that powerful models can facilitate real cyberattacks, a specialized version of Large 4 for advanced security tasks is only available to corporate partners under strict controls.
Testing partners include Booz Allen Hamilton and CrowdStrike, while Cisco and Cloudflare have provided support for the model's release. The European Union’s cybersecurity agency, ENISA, has tested Large 4 for vulnerability detection and code audit.
Large 4 is distributed under the Mistral Ecosystem Open-Weight license, a modified MIT license, allowing for free use, modification, retraining, and integration into commercial products.
However, companies seeking to sell access to Large 4 as a cloud service must negotiate separately with Mistral. The data used to train the model remains proprietary.
Beyond its own servers, users can run Large 4 through Mistral Studio, API, and OpenRouter. Pricing is set at $0.70 per million input tokens, $0.07 for cached tokens, and $2.10 for output tokens.
Notably, the public preview of Large 4 was initially launched on October 6, following training on 3,800 Nvidia Grace Blackwell accelerators in European data centers.
