Mistral AI Previews Mistral Large 4, a 1 Trillion Parameter Model With Weights Due in October

French AI company Mistral AI on October 6, 2026 launched a public preview of Mistral Large 4, a natively multimodal model with about 1 trillion parameters that it calls its largest and most capable to date. The API is available now, and Mistral says it will release the model weights by the end of October.

Mistral nicknames the model "le Chonk" in its announcement and positions it as an open weight alternative to closed US models and to the open models from Chinese labs that lead most rankings. It is the first model on the roadmap funded by a €3 billion Series D, which Mistral calls the largest equity round ever raised by a European technology company; Samsung led that round in September 2026 at a €21 billion valuation, about $24.4 billion (TechCrunch).

What did Mistral announce?

Mistral Large 4 is a mixture of experts model, a design in which only a small share of the network runs for each token, with text and image input and text output. In its announcement, Mistral gives 1 trillion total parameters and 49 billion active per token; its model documentation lists 1.05 trillion total, 52 billion active and a 1.6 billion parameter vision encoder.

Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and that the preview runs on the same hardware. More than 160 languages appear in the training data, including every official language of the European Union, according to the company. Pierre Stock, Mistral's VP of Science, told TechCrunch the model was trained on about 4,000 Nvidia GPUs, "two to three times less than our Chinese competitors"; the announcement gives the exact count of 3,800.

The model combines instruction following and reasoning in one system. Developer Simon Willison reports that the API exposes only two reasoning levels, "none" and "high". Mistral also says the reinforcement learning run behind the preview is still going and that it expects "large and rapid improvements" in the coming weeks.

When will the weights be released?

Mistral says it will publish the weights of Mistral Large 4 by the end of October 2026, together with details on the architecture, more benchmarks and its post-training method. Until then, the model is available only through APIs.

During this period, Mistral says it is red teaming the model with cybersecurity firms, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities. Stock explained the delay to TechCrunch:

Until then the model runs on Mistral Studio and is listed on Vercel AI Gateway and OpenRouter. Mistral is not alone in announcing weights before shipping them: Reflection AI promised the weights of its 501 billion parameter Beam for later in October a day earlier, while Aleph Alpha's Kolibri shipped with its weights on October 3.

How does Mistral Large 4 compare with other models?

Independent evaluator Artificial Analysis scores the Mistral Large 4 preview at 38 on its Intelligence Index, which it says makes it the most capable model from outside the US and China. According to The Decoder, Mistral Large 3 scored 9 on the same index, while Claude Opus 5.5 (Max) leads it with 58 points, 20 more than Mistral Large 4.

Artificial Analysis also flags cost: even at the launch discount, a task on its index costs more than twice as much with Mistral Large 4 as with GLM-5.3-Flash or DeepSeek V4.1 Flash. Mistral's own coding numbers include 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, which the company reports and which have not been reproduced independently.

Why is Mistral focusing on cybersecurity?

Mistral argues that open weights matter most in cybersecurity, where closed models can refuse legitimate vulnerability research. It says Mistral Large 4 scores 82% on a test that asks a model to reproduce and then patch a real flaw in open source software, while Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.

The Decoder points out that this result measures provider policies as much as model ability. Mistral also says the model refuses malicious cyber prompts more often than any other open model on its tests, without explaining how it separates defense work from attack preparation. Anthropic went the other way on the same day, putting its least restricted cyber models behind verified access tiers for vetted defenders.

How much does it cost?

Mistral lists standard API prices of $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. Artificial Analysis reports a 50% discount for the first two weeks, which brings the rates to $0.68, $0.07 and $2.09.

What we don't know yet

• Which license will cover the weights, and whether it allows unrestricted commercial use.
• Which parameter count is final: the announcement and the documentation give different figures.
• Whether the full 1 million token context will be available on the API and in the released weights.
• What hardware is needed to run the full model, and whether smaller variants will follow.
• How the released weights will score once the reinforcement learning run ends.