Mistral Large 4 enters public preview, with open weights due by the end of October

Mistral AI launched a public preview of Mistral Large 4, its largest model to date, on October 6, 2026. The company gives it the unofficial short name ML4 and, with a wink, the nickname "le Chonk". For now the model can only be tried through the API on Mistral Studio, the company's developer console. According to Mistral's announcement, the open weights will be released by the end of the month; as of today, they are not yet available for download.

Mistral Large 4 is a general-purpose language model that combines instruction following, step-by-step reasoning and agentic abilities, meaning it can use tools to complete tasks, in a single model. Alongside text, it accepts images as input, including documents, charts, photos, engineering drawings and satellite imagery. Mistral is positioning it for enterprise work such as software development, cybersecurity, finance and law. Once the weights are out, Large 4 will also be an open model you can run on your own servers.

A trillion-parameter mixture-of-experts model

According to Mistral's documentation, the model has 1.05 trillion parameters in total and uses a mixture-of-experts (MoE) architecture. In this design the model is made of many expert subnetworks, and only some of them are used for each token; in Large 4, 52 billion parameters are active at each step. A separate vision encoder that reads images has 1.6 billion parameters. The documentation lists a context window, the amount of text the model can read at once, of 1 million tokens.

According to the company, the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, and the preview is served on the same infrastructure. Mistral says a significant share of the training data was multilingual, covering more than 160 languages, including every official language of the European Union. Reinforcement learning on the model is still running, and the company says it expects clear improvements in the coming weeks.

Benchmark results shared by the company

According to Mistral's published results, the model scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. The company says that on a combined coding agent index it places ahead of DeepSeek V4 Pro and Qwen3.8 Max. In a blind human evaluation of coding quality run with Surge AI, with model names hidden, it ranked second of five models, behind Claude Opus 5.

Cybersecurity gets the most attention in the announcement. According to Mistral, the model ranks among the top five worldwide on the Artificial Analysis Cyber Index. It scored 82% on a test that asks a model to reproduce a real vulnerability in open-source software and then patch it, and 93% on Cybench, a set of challenges drawn from security competitions. The company says some closed models score very low on the same test because they refuse the task. According to Mistral's measurements, the model also refuses malicious cybersecurity requests more often than the other open models it was compared with, and it resists 93.3% of attacks on Lakera's B3 security benchmark.

On vision, the company reports that the model edges out GPT-6-Astra, 42% to 41%, on Dense 200, a test of finding and marking objects in an image. In finance and legal evaluations that Mistral had run by vals.ai, the model is also reported to come out ahead of GPT-6-Astra.

Access, weights and next steps

The preview API can be called on Mistral Studio as mistral-large-4. The documentation lists features such as structured outputs, function calling, batch processing and an agents interface. Current prices are on Mistral's site.

Mistral says it will release the weights by the end of the month, together with more details on the architecture, additional benchmarks and its post-training methods. Until then, the model is being red-teamed in real-world settings with cybersecurity firms, vetted partners and state authorities, who get access to the same model with reduced moderation and expanded cyber capabilities. The company says the model will be available in several regions worldwide, including a European deployment that Mistral operates end to end. According to the announcement, it will also be able to run in a private cloud or on premises. Mistral adds that Large 4 will serve as the base for a new generation of specialized models for specific industries.

The model is still in preview, and since training is ongoing, its capabilities and scores are expected to change. Details of the architecture have not been released yet. The announcement does not say which license the weights will use, so anyone planning commercial use will need to check the terms once the files are out. Running a model of this size yourself will take several data center GPUs; the company has not yet shared hardware requirements. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Mistral AI: Introducing Mistral Large 4, Mistral Docs: Mistral Large 4.