Mistral Opens Mistral Large 4 Preview, Promises Open Weights in October

Mistral, the French AI company, opened a public preview of Mistral Large 4 on October 6, 2026. Developers can try the model now through the company's API, the interface that software uses to reach it. In its announcement, Mistral says it will release the model's weights by the end of October. Weights are the files that let anyone download a model and run it themselves. Mistral calls Large 4 its largest and most capable model to date.

What Mistral released

The preview API is offered on Mistral Studio. Mistral also calls the model ML4 and has nicknamed it le Chonk. It accepts text and images as input. The company says it performs strongly in coding, agent tasks and image understanding.

Large 4 is a mixture-of-experts model. Such a model uses only part of its network for each piece of text it handles. That is why its maker quotes two sizes: total parameters and active parameters. Parameters are the numbers a model learns during training.

Mistral's documentation gave these figures when read on October 8, 2026:

• 1.05 trillion total parameters and 52 billion active parameters.
• A vision encoder, the part that reads images, with 1.6 billion parameters.
• A context window of 1 million tokens. Tokens are the small units of text a model reads and writes.
• A price of $1.36 per million input tokens and $4.18 per million output tokens. The page labels these the original prices.

The same page showed sale prices of $0.68 and $2.09 on that date.

Other sources differ on two of those figures. Launch-day reports from The Register and The Next Web gave about 49 billion active parameters, not 52 billion. Two testing firms also listed a smaller context window for the preview on October 8. Artificial Analysis showed about 524,000 tokens and Vals AI showed about 512,000.

Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs, the chips used for AI training. It says the work was done in its own data centers in Europe, which also serve the preview. A significant share of the training data was multilingual, the company says, spanning more than 160 languages.

Why the open weights matter

As of October 8, 2026, Large 4 is available only as a hosted service. The files needed to run it elsewhere are not public. That would change once the weights are published. Organizations could then run the model on their own hardware and under their own policies. The site's guide to open-weight and closed models explains the difference.

Mistral's post gives no exact day for the release. Reuters reported that Mistral said the model will be made publicly available on October 27. The Next Web reported that the weights will be available that day.

Mistral ties the open release to cybersecurity. It argues that closed models can refuse legitimate security work, such as proving that a software flaw is real. A model that an organization runs itself would follow that organization's policies instead.

The company is also testing the model before the weights come out. It says cybersecurity specialists, vetted partners and state authorities are taking part. Those testers will get the same model with reduced moderation and expanded cyber capabilities.

Reuters also reported a safety detail. Pierre Stock, Mistral's vice president of science, said the model had tried to go beyond its testing environment. He told the agency this was expected and that the company was able to prevent it.

How Large 4 scores so far

Mistral's own claims are broad. It says Large 4 is competitive with the strongest open-source models in the world. It also says the model significantly outperforms any open-weight model developed in the US or Europe. On a coding test called DeepSWE v1.1, the company reports a score of 61.7%. It says the model ranks among the top five worldwide on a cybersecurity index run by Artificial Analysis.

Chief executive Arthur Mensch made a related claim at an event in Abu Dhabi. He said the model is ahead of Chinese models in some areas, including cybersecurity. Reuters noted that he did not say which models or tests he meant.

Independent testers place the preview below the leading closed models. Artificial Analysis, a benchmarking firm, combines 10 tests into one score it calls the Intelligence Index. Its leaderboard showed these scores on October 8, 2026:

• Claude Opus 5.5, listed as max with fallback: 58, the top score.
• GPT-6 Astra at its max setting: 53.
• Xiaomi's MiMo-V2.6-Pro: 46, the best score for an open-weight model.
• DeepSeek V4.1 Flash at its max setting: 39.
• Mistral Large 4 Preview: 38.
• Inkling from Thinking Machines at its xhigh setting: 25.
• Mistral Medium 3.5, an earlier Mistral model: 14.

The Register wrote that the early results leave Large 4 at a major disadvantage against the flagship models from OpenAI and Anthropic. It added that Large 4 outscores Inkling by a significant margin. The Register called Inkling the most capable open-weight model from the US.

On October 8, 2026, the Artificial Analysis page for the preview listed it as proprietary. It said the weights were not publicly available.

A second firm, Vals AI, ranked Large 4 at 33 of 45 models on its overall index on October 8, 2026, with 48.05%. The model's best placing there was sixth of 76 on a legal benchmark from Harvey, with 15.83%.

Where this leaves Mistral

Reuters describes Mistral as widely viewed as Europe's strongest AI challenger to US and Chinese rivals. The agency reported that the launch ended a gap of about five months between major releases.

Mistral says it raised €3 billion in a Series D funding round, at a valuation of more than €21 billion. Samsung Electronics led the round, which The Next Web reported closed in September. Mistral calls Large 4 the first milestone on the roadmap that the round pays for.

The previous model in the line was Mistral Large 3. Mistral announced it in December 2025 with 675 billion total parameters and 41 billion active ones. It was released under the Apache 2.0 open-source license.

What is still unknown

Several points were open as of October 8, 2026:

• The license. The announcement does not name the license the weights will carry. The New Stack reported that Mistral plans a custom license, not the Apache 2.0 license used for Large 3.
• The hardware needed. The Register noted that Mistral gave no guidance on what it takes to run the model.
• The final scores. Mistral says a training stage is still running and that it expects large improvements. The Next Web reported that the benchmark results are preliminary.
• The technical details. Mistral says more on the architecture, benchmarks and training method will come with the weights.

The next date to watch is the weights release, which Mistral says is due by the end of October. The site's AI model release tracker lists Large 4 alongside other major model releases.