Mistral Large 4 Preview Brings 1 Trillion Open-Weight Model

Mistral Large 4 launched in preview Oct. 6, 2026: about 1 trillion parameters, multimodal, with open weights promised by the end of October.

Updated

Share this article

Home › Tech, AI and Business

Mistral Large 4 is here in preview. The Paris-based AI company launched a public preview on Tuesday, Oct. 6, 2026, calling it its largest and most capable model to date: a natively multimodal system with about 1 trillion parameters that developers can try now through the Mistral Studio API, with open weights promised by the end of October, according to Mistral’s announcement.

Part of our Tech, AI and Business coverage. Browse more in Artificial Intelligence.

Advertisement

Mistral Large 4 specs at a glance

Mistral, which unofficially calls the model ML4 and “le Chonk,” uses a mixture-of-experts design, so only a slice of its parameters is active for any given token. The company’s model documentation lists 1.05 trillion total parameters, 52 billion active parameters including embedding and output layers (49 billion otherwise), a 1.6 billion-parameter vision encoder and a context window of 1 million tokens. It supports structured outputs, function calling, document question answering, batching and Mistral’s agents tools. The documentation lists preview pricing of $0.68 per million input tokens and $2.09 per million output tokens. Mistral says a significant share of the training data was multilingual, covering more than 160 languages, including every official EU language.

Trained in Europe on Nvidia Grace Blackwell chips

Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and that the public preview is served from that same infrastructure. The Next Web reported that training took about two months and roughly 10 megawatts of power, and that Mistral is pitching Large 4 as the most powerful open-weight AI system built outside China. That framing matters in Europe, where governments want homegrown AI; Germany’s Aleph Alpha made a similar open-weight bet, as we covered in our report on Aleph Alpha Kolibri.

What Mistral says Large 4 can do

Every performance claim below comes from Mistral’s own announcement and has not been independently verified by The Boston Report. Mistral says the model is state-of-the-art among open models on cybersecurity, finance and law workloads, and that it surpasses even some frontier closed models at visual grounding.

  • Cybersecurity: Mistral says ML4 ranks among the top five models on the Artificial Analysis Cyber Index and scores 82% on one test that asks a model to reproduce a real open-source vulnerability and then patch it. It also says the model solves 93% of the challenges in Cybench, a set of 40 security-competition exercises.
  • Coding: Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, and a combined Coding Agent Index of 49.8%.
  • Human ratings: In a blind evaluation by Surge AI, professional annotators ranked the preview second of five models for coding quality, at 3.74 on a 1-to-5 scale, behind only Claude Opus 5 at 4.22.

When the open weights arrive

Mistral says it will release the weights “by the end of the month,” along with more detail on the architecture, additional benchmarks and its post-training method. The Next Web reported a target of Oct. 27, while the model’s Hugging Face page shows an expected release date of Oct. 31. Until then, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.

How Mistral Large 4 fits the AI race

The release lands in a crowded month. Google has been rolling out Gemini 4 Argon to cyber defenders first, and OpenAI pulled back a model over safety concerns, as in our story on OpenAI GPT-6.1 Astra. Mistral’s angle is openness: once the weights ship, companies will be able to run a trillion-parameter model on their own servers, and Mistral says Large 4 will also be the base for a new set of specialized models. Independent testing after the weights are public will show how the self-reported numbers hold up.

Related: FLUX 3 Image Brings Bounding-Box Layout and 4K Edits

Sources

Further Reading

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up