Mistral Large 4 has one trillion parameters, and Mistral says it will release them as open weights before the end of October. The French AI company opened a public preview of its new flagship on 6 October and said the model's weights, the trained numbers that make it work, will be released this month, giving companies and developers a frontier-class model they can host, inspect and modify themselves.
That matters beyond the benchmark tables. Most of the strongest AI systems today are closed, rented through an API on the provider's terms. An open-weight model of this size, built and trained in Europe, gives builders another option at a moment when where AI runs, and who controls it, has become a business decision in its own right.
What Mistral Released and When
Mistral calls the model ML4 internally and, in its own announcement, "very officially: le Chonk." The preview API is available now through Mistral Studio, and the company says it "will release the weights by the end of the month." Until then, Mistral says it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who are testing a version with reduced moderation and expanded cyber capabilities.
The company describes ML4 as a natively multimodal model with one trillion parameters in total and 52 billion active parameters, meaning only a fraction of the network is used to process any given request. That design helps make a model of this scale practical to serve. Mistral Large 4 is the company's largest and most capable model to date, and Mistral says it is still improving as training continues.
More detail is promised before the weights arrive. Mistral says it will publish further information on the model's architecture, additional benchmarks and its post-training methodology, and that ML4 will become the foundation for a new generation of specialised and optimised Mistral models.
Built in Europe, on Its Own Hardware
The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centres in Europe, and the public preview runs on the same infrastructure, according to the company. CNBC reported a figure of roughly 4,000 GPUs over two months of training.
Language coverage is broad. A significant share of the training data was multilingual, spanning more than 160 languages including every official language of the European Union, which matters for businesses and public bodies that work outside English. ML4 will be offered in multiple regions worldwide, including a European deployment that Mistral says it operates end to end, independently of other digital service providers and under European law.
Chief executive Arthur Mensch framed that independence as the point. Speaking to CNN, Mensch said the European build keeps Mistral outside a US-China "duopoly" in AI. "There's an enormous desire and an enormous need for alternative technology suppliers," Mensch added.
What the Benchmarks Claim
Mistral's published results put Mistral Large 4 at or near the top of open-weight models on several tasks. On software engineering it reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, for a combined Coding Agent Index score of 49.8%, which the company says places it ahead of DeepSeek V4 Pro and Qwen3.8 Max. In a blind human evaluation of coding quality run with Surge AI, ML4 ranked second of five models, behind only Claude Opus 5.
On business automation, Mistral reports 59.9% on AutomationBench, a set of 657 workflows across tools such as Gmail, Google Sheets, Slack and Salesforce, and a score of 1,393 Elo on AA-Briefcase, which tests long knowledge-work tasks such as spreadsheets, slides and PDFs. On visual grounding, it says ML4 edged GPT-6 Astra on the Dense 200 test, 42% to 41%.
These are the company's own figures, and they deserve context. CNBC noted that the model still lags the frontier in areas such as coding, and CNN said it could not immediately verify Mistral's cost and capability claims. Independent testing will follow once the weights are public.
Why Cyber Defence Is the Headline Use
Mistral is leading with security. On the Artificial Analysis Cyber Index, an independent evaluation of how well models find and fix flaws in real software, the company says ML4 ranks among the top five models globally and leads open-weight models developed outside China. On one test that asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scored 82%, which Mistral says is the highest of any model, and it solved 93% of Cybench, a set of 40 security-competition exercises.
The company's argument is about refusals. It says several leading closed models score near zero on that vulnerability test because they decline the task, even though proving that a flaw is real is often the first step in defending against it. An open-weight model running on an organisation's own systems avoids that problem, and keeps sensitive data in-house. Co-founder and chief scientist Guillaume Lample told CNBC that the model's cyber defence capabilities "will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks."
Other companies have made the same case from experience. Hugging Face chief executive Clement Delangue told CNN in July that, during a security incident at the company, running a model on its own infrastructure was mandatory because the data involved was private, and that only open models made that possible.
What Open Weights Mean for Smaller Builders
For startups, small agencies and independent developers, open weights change the bargaining position. A team can fine-tune the model on its own data, run it in its own cloud account or in a region of its choosing, and avoid having a product depend on a single vendor's pricing or policy decisions. Mistral's VP of science, Pierre Stock, told CNN the model was built at a "fraction of the compute" of competitors, which the company says makes it cheaper to build and to use.
There is a practical limit. A trillion-parameter model is not something most small teams will run on a laptop or a single server; serving it takes serious hardware. For many builders the realistic path will be the hosted API, a cloud provider that offers the open weights, or the smaller specialised models Mistral says it will derive from ML4. The point of open weights is that those options exist and compete, rather than that every team runs the full model itself.
At MW3.biz, we believe capable AI should be available to as many builders as possible, not only to the largest companies. From a technology-democratisation perspective, a frontier-class model released with open weights widens the set of people who can study, adapt and deploy advanced AI. Concerns about misuse are real and deserve careful handling, and Mistral's staged red-teaming is one response to them. Our view is that broader, accountable access remains the better long-term answer.
Where Mistral Goes Next
The company has the money to keep scaling. Mistral raised a 3 billion euro Series D in September at a 21 billion euro valuation, according to CNBC, and Lample said model capabilities will improve further as training capacity grows on the back of that round. CNN put Mistral's total funding at about $6 billion since its founding three years ago, a fraction of what the largest US labs have raised.
The competition is crowded. Chinese labs including DeepSeek, Alibaba and Z.ai lead much of the open-weight market, and American startup Reflection released its own open model, Beam, the day before Mistral's announcement. For builders comparing options, last week's coverage of OpenAI Codex running in the cloud shows how quickly closed coding tools are moving too.
The next milestone is the weights themselves. Once Mistral Large 4 is downloadable, independent benchmarks, licence terms and the hardware needed to serve it will show how far an open trillion-parameter model really travels beyond Mistral's own infrastructure.
