Mistral Opens the Weights of Its Trillion-Parameter Large 4 Model, Betting Open AI Can Win on Coding and Cyber Defense

PARIS — French artificial intelligence lab Mistral AI has released the open weights of Mistral Large 4, its trillion-parameter flagship model nicknamed "Le Chonk," giving developers, companies and researchers the freedom to download, inspect, modify and run the system on their own servers. The release, announced on Thursday, October 29, 2026, turns a model that had only been available through Mistral's own preview service since October 6 into one of the largest open-weight systems ever published.

Mistral is pitching the model first and foremost at two jobs: writing and improving software, and defending computer networks. The company says Large 4 is its strongest open model yet for agentic coding — AI systems that can carry out multi-step programming tasks — and for defensive cybersecurity work, a combination it hopes will set it apart in a crowded field of open models led by Chinese rivals.

The release matters beyond a single product launch. It is the most ambitious statement yet from Europe's leading AI lab that powerful models need not stay locked inside the companies that train them, and it arrives at a moment when investors are asking hard questions about how much the AI boom costs and who ultimately pays for it.

Illustration: A modern AI data center hall with rows of server racks, representing the infrastructure behind a large open-weight model.

What Happened

Mistral published the weights of an improved version of Large 4 on Hugging Face, the platform where developers share and download AI models, on October 29, 2026. Weights are the numerical values a trained model has learned; releasing them means anyone with sufficient computing hardware can run the model themselves, fine-tune it for a particular task, or study how it behaves, rather than renting access through an API alone.

According to Mistral's model documentation, Large 4 is a mixture-of-experts system with roughly one trillion total parameters — the internal values that shape its responses — of which roughly 50 billion are active for any given token of text. It accepts text and images as input, produces text and code, and supports a context window of one million tokens, meaning it can work with very long documents or codebases in a single session. Mistral also says the model is natively fluent in more than 160 languages.

The released version is not identical to the October 6 preview. Mistral says it reflects three additional weeks of reinforcement learning — training that rewards better answers — and testing with customers, and that it improves on the preview across coding, defensive cybersecurity and reasoning efficiency. The company says it trained the model from scratch in its own data centres in Europe, on thousands of Nvidia Grace Blackwell graphics processors.

Alongside the weights, Mistral continues to offer the model through its own API and Mistral Studio platform, and through partners including Microsoft Foundry, which Mistral describes as the first cloud to carry it, and Nvidia, which offers it as a NIM container. API pricing is listed at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

Illustration: Developers working with AI-assisted coding tools in an office, representing the programming tasks Mistral Large 4 is built for.

Background

Open-weight models occupy a middle ground in the AI industry. Unlike fully open-source software, their training data and code are not necessarily public, but their weights are — enough for others to run, adapt and build on them without permission from the original developer. That model has powered a fast-moving ecosystem of start-ups, researchers and hobbyists, and it has become a strategic question for governments: whoever controls the most capable open models influences what smaller companies and public institutions can build without depending on a handful of American or Chinese providers.

Mistral, founded in Paris in 2023, has made openness and European sovereignty central to its identity. Large 4 is its flagship, and the company had promised at the October 6 preview that the weights would follow by the end of the month. Keeping that promise on October 29, after a period of red-teaming with cybersecurity leaders, vetted partners and state authorities, was itself part of the message: the model, Mistral argues, can be released openly because its safeguards were tested before publication.

The terms of the release reflect that balancing act. Large 4 is published under the Mistral Ecosystem Open-Weight License, a modified MIT licence that permits free use, modification, fine-tuning and commercialisation — with one notable condition. Companies that want to offer Large 4 itself as a hosted model service need a separate commercial licence from Mistral. Ordinary businesses building products on top of the model are not caught by that condition, but cloud providers reselling the model as a service are.

Self-hosting is not trivial. Published deployment guidance indicates the standard version needs eight Nvidia B300 graphics processors, with a more compressed variant running on four GB300 processors — hardware far beyond a single desktop computer, but within reach of a well-equipped company data centre. That is the point of a trillion-parameter model: it is designed for organisations, not individuals.

Why It Matters

For developers, the headline numbers are about coding. Mistral reports that Large 4 scores 19.7 percent on Cognition's private FrontierCode 2.0 benchmark, the top score among open-weight models, and 48 percent on Artificial Analysis's Coding Agent Index — effectively tied with China's GLM-5.3, but using 47 percent fewer tokens per task, which translates directly into lower running costs. Its score on Terminal-Bench 4.0, a test of real-world terminal-based programming tasks, rose to 41 percent, up from 28 percent at preview. All of these figures come from Mistral itself or from the benchmark providers it cites, and independent validation will take time.

For security teams, the pitch is defensive. Mistral says Large 4 scores 51 percent on the Artificial Analysis Cyber Index and that, in its tests, it refused 96.6 percent of malicious cyber requests on average — more than any other open model, according to the company. At the same time, Mistral acknowledges the dual-use tension at the heart of cybersecurity AI: a model good at finding vulnerabilities could also make real attacks easier. A more aggressive penetration-testing version, which can actively probe systems for weaknesses, is therefore not in the open release at all; it is restricted to vetted partners under contractual conditions.

For Europe, the release is a sovereignty play. A model trained in Mistral's own European data centres, that organisations can inspect and run on their own infrastructure, offers governments and regulated industries an alternative to sending sensitive data to foreign-owned services. That argument has grown louder as AI systems move from chatbots into core business operations, where control over data and models becomes a board-level concern.

And for the wider AI market, the timing is striking. The release landed in the same week that investors punished big technology stocks over the cost of the AI build-out, with Alphabet shares suffering a record fall after its cloud spending pushed quarterly cash flow negative. Against that backdrop, an open model that claims to do comparable coding work with far fewer tokens per task is not just a technical claim — it is an economic one.

Illustration: A digital shield protecting a computer network, representing the defensive cybersecurity role Mistral pitches for Large 4.

What Happens Next

The immediate test is adoption. Developers will now try Large 4 on their own codebases and security workloads, and independent benchmarks in the coming weeks will show whether Mistral's coding and cyber claims hold up outside its own tests. Independent scoring already paints a more mixed picture on general intelligence: Artificial Analysis's Intelligence Index rates Large 4 at 26 points, below DeepSeek V4.1 Flash and GLM-5.3 at 32, a reminder that a model can lead on coding efficiency while trailing on broader capability.

Competition will not stand still. Moonshot AI's Kimi K3.1 Pro, whose weights are due by November 4, 2026 according to reports, tops Terminal-Bench 4.0 at 67 percent in Mistral's own comparison notes — though its weights are not yet public. The open-weight race, in other words, is still being run, and Large 4's lead is specific to the coding and defensive-security tasks Mistral chose to emphasise.

Regulators and enterprise buyers will also watch how the licence condition plays out in practice: where building a product on Large 4 ends and reselling Large 4 as a service begins is a boundary that lawyers, not engineers, will have to police. And Mistral says it plans to tighten safeguards further in the coming months, suggesting the safety debate around powerful open models is far from settled.

For now, though, the facts are simple. Europe's flagship AI lab has put a trillion-parameter model into the hands of anyone with the hardware to run it — and bet that openness, efficiency and defensive security are enough to win developers over.

Sponsored: explore today's featured sponsored offers

FAQ

What is Mistral Large 4?

Mistral Large 4, nicknamed "Le Chonk," is Mistral AI's flagship AI model, with roughly one trillion total parameters. It handles text and image input and produces text and code, and it is designed for agentic coding, cybersecurity and enterprise AI agents.

What does releasing the "open weights" mean?

It means the trained model's numerical weights are published, in this case on Hugging Face, so developers and organisations can download and run the model on their own servers, fine-tune it, and study its behaviour, instead of only accessing it through Mistral's API.

Is Mistral Large 4 free to use?

Use, modification, fine-tuning and commercialisation are permitted under Mistral's modified MIT open-weight licence. However, businesses that offer Large 4 itself as a hosted model service need a separate commercial licence from Mistral. API access through Mistral is priced per token.

How good is Large 4 at coding?

According to Mistral, it scores 19.7 percent on FrontierCode 2.0, the top open-weight score, and 48 percent on the Coding Agent Index, tied with GLM-5.3 while using 47 percent fewer tokens per task. These are company-reported figures that independent testers will now verify.

Why is Mistral restricting the penetration-testing version?

Mistral says a model skilled at finding vulnerabilities could also make real attacks easier. The more aggressive penetration-testing version is therefore available only to vetted partners under contractual restrictions, while the open release is aimed at defensive cybersecurity work.

Sources

• Mistral AI announcement of the Mistral Large 4 open-weights release, October 29, 2026, as reported by Skimwire, Unite.AI and TestingCatalog.
• Dev.to, "AI Daily Digest October 30, 2026: Mistral Opens a Trillion Parameters, Cloud Giants Report, OpenAI Rewrites the Spec," October 30, 2026.
• Pivot News, "Mistral ships Large 4 weights, strong at coding but behind China..." October 30, 2026, including Artificial Analysis Intelligence Index scores and self-hosting hardware requirements.
• Reuters, "Stocks slump, bonds leap as AI giants' spending spooks investors," October 30, 2026, for market context on AI spending concerns.