With Mistral Large 4, French AI company Mistral AI unveiled its largest and most capable model to date on 6 October 2026. Internally, it carries the rather unapologetic nickname “Le Chonk”. But there is more behind the unusual name than sheer scale: with its latest model, Mistral is attempting to narrow the gap between European open-weight systems and the most capable models emerging from the United States and China. Mistral Large 4 has been available as a public preview through Mistral Studio and the API since this week, with the model weights expected to be released towards the end of October.
The timing is strategically significant. While OpenAI, Anthropic and Google continue to offer their most powerful systems largely as closed platforms, and Chinese companies have recently made considerable progress with open models, Mistral is once again betting on a combination of high performance, European infrastructure and the ability for organisations to eventually run the model themselves. Large 4 is therefore intended to be more than simply the company’s next large language model. It is designed to become the technical foundation for a new generation of specialised Mistral systems.
One Trillion Parameters, but Only a Fraction Are Active at Any Time
Mistral Large 4 is a Mixture-of-Experts model. In total, the system contains around 1.05 trillion parameters, although only a relatively small proportion is activated for each request. Mistral’s official announcement cites 49 billion active parameters, while parts of the current documentation give a slightly different figure of 52 billion. The model also incorporates a 1.6-billion-parameter vision encoder.
For practical use, this architecture is crucial. It allows Mistral to build an exceptionally large model without having to process every parameter for every token, potentially improving the balance between capability and computational efficiency.
The model is natively multimodal and can process text and images together. It also combines instruction-following, reasoning and agentic capabilities within a single system. In doing so, Mistral is attempting to reduce the growing separation between fast conversational models, dedicated reasoning models and specialised agent models.
Large 4 also offers a context window of up to one million tokens. This makes it particularly relevant for tasks involving large codebases, extensive document collections, technical archives or complex corporate information. It supports structured outputs, function calling, document question answering and batching, as well as Mistral’s Agents and Conversations interfaces.
More Than 160 Languages and a Distinctly European Ambition
Mistral places particular emphasis on multilingual performance. A significant proportion of the training data is said to consist of non-English material. According to the company, Large 4 supports more than 160 languages, including every official language of the European Union.
Its European positioning extends well beyond language support. Mistral says the model was trained from scratch in Europe using 3,800 NVIDIA Grace Blackwell GPUs in the company’s own European data centres. The current public preview is also being hosted on this infrastructure.
This is a central part of Mistral’s broader strategy. In the long term, companies and public institutions are meant to gain access not only to a European AI model, but also to greater control over where their AI systems are operated and which technological dependencies they accept.
The Real Focus Is on Agents
Large 4 becomes particularly interesting when it comes to agentic tasks. Mistral is positioning the model less as a conventional chatbot and more as a foundation for systems capable of using tools, gathering information and carrying out complete workflows.
On AutomationBench, a benchmark covering hundreds of business workflows across applications such as Gmail, Google Sheets, Slack and Salesforce, the preview reportedly scores 59.9 per cent. According to Mistral’s comparison data, its predecessor achieved just 6.3 per cent. The scale of that increase illustrates where much of the company’s development effort has been directed: away from pure question answering and towards practical tool use and process automation.
Mistral is also emphasising the creation of professional deliverables. Large 4 is designed to handle extended chains of tasks and turn them into finished spreadsheets, presentations, PDFs and other work products. That places it squarely in the same market in which OpenAI, Anthropic and other providers are increasingly positioning their models as digital work agents.
Strong at Coding, but Not the Undisputed Leader
Large 4 also represents a significant step forward in programming. The model achieves 61.7 per cent on DeepSWE v1.1, 59.4 per cent on SWE-Atlas-QnA and around 28 per cent on Terminal-Bench 4.0. In a blind evaluation of professional coding responses conducted by Surge AI, Large 4 reportedly ranked behind Claude Opus 5 but ahead of several current open competitors.
This is also where the limitations of the initial announcement become apparent. Mistral Large 4 is clearly a capable coding model, but it does not consistently lead the overall market. German technology publications have also noted that leading systems from OpenAI and Anthropic continue to perform considerably better on some programming benchmarks. Many of Mistral’s own comparisons also focus primarily on open-weight competitors.
That does not make Large 4 any less interesting, but it changes the interpretation. Mistral has not suddenly overtaken every frontier model. Instead, it has produced a highly capable European open-weight model with a distinctive set of strengths.
Cybersecurity Becomes Its Most Unusual Selling Point
Mistral is positioning Large 4 particularly aggressively in cybersecurity.
On CyberGym-E2E, a benchmark in which models must reproduce and then fix real vulnerabilities in open-source software, Large 4 reportedly achieves 82 per cent. On Cybench, it completes 93 per cent of the 40 challenges. Mistral also highlights a strong position on Artificial Analysis’s Cyber Index.
The company is deliberately taking a somewhat different approach from certain closed-model providers. Mistral argues that security researchers need capable models that can genuinely understand and reproduce real vulnerabilities. Safety filters that are too restrictive, it suggests, can interfere with legitimate defensive work.
Large 4 is therefore able to perform certain security-related tasks that other models may refuse because of their safety policies. That creates an obvious tension. Powerful cyber capabilities can be extremely valuable to defenders, but the same abilities can theoretically be misused for offensive purposes.
For this reason, Mistral is using the period before the model weights are released for additional red-team testing with cybersecurity firms, selected partners and government organisations. Some of these groups are reportedly being given access to a less heavily moderated version in order to assess the model’s true risk profile.
Multimodality Matters More Than Another Chat Feature
Another major improvement lies in visual understanding.
Large 4 is designed to analyse complex documents, diagrams, technical drawings and natural images. Mistral demonstrates this capability with examples such as satellite imagery and engineering drawings, where the model can inspect specific regions and connect visual information with other contextual data.
For industries such as manufacturing, engineering, logistics, geospatial analysis and scientific research, this combination of visual understanding and agentic capabilities may prove far more significant than conventional image description. The model is intended not merely to recognise visual information, but to use it as part of longer decision-making and problem-solving workflows.
That fits closely with Mistral’s broader strategy of targeting enterprise and industrial customers.
A One-Million-Token Context Window Expands the Possible Use Cases
The one-million-token context window is particularly relevant for professional workloads.
Instead of uploading individual documents one by one, organisations can potentially work with very large bodies of information within a single context. Possible use cases include extensive contract archives, technical documentation, large software repositories and research collections.
A large context window alone does not guarantee that every detail will be handled reliably. What matters is how effectively the model can retrieve and connect relevant information across such a large volume of material. Independent testing after the full release will therefore be particularly important for assessing Large 4’s real long-context performance.
The Introductory Pricing Is Strikingly Aggressive
Mistral is also attempting to put pressure on competitors through pricing.
During the current introductory period, one million input tokens cost $0.68, cached input costs $0.07 and one million output tokens cost $2.09. However, these are temporary launch prices reflecting a 50 per cent discount. The regular list prices are $1.36 per million input tokens and $4.18 per million output tokens.
Even at the standard rate, that is an aggressive price for a model of this scale. It fits Mistral’s established strategy of competing not purely on benchmark leadership, but on the overall balance between performance, cost, openness and deployment flexibility.
Open Weight Is Not the Same as Open Source
A central part of the announcement is the planned release of the model weights at the end of October.
This should allow companies to run Large 4 on their own infrastructure or within private cloud environments. For sensitive data, cybersecurity, public-sector use and regulated industries, that could be strategically important.
However, the terminology matters. Open weight means that the trained model parameters are made available. It does not automatically mean that the training data, training code and complete development process are also open.
Even so, Mistral’s approach offers substantially greater technical control than models available exclusively through proprietary cloud APIs.
Europe’s Answer to American and Chinese AI
The significance of Large 4 therefore extends beyond its benchmark scores.
Chinese providers including DeepSeek, Alibaba, Z.ai and Moonshot AI have made substantial progress in open models, while the most powerful closed systems continue to come predominantly from American companies.
Mistral is attempting to establish a European alternative between those two poles. Large 4 is intended to be capable enough for demanding enterprise workloads while remaining open enough for organisations to run, adapt and control independently.
Independent rankings also suggest that Large 4 is not automatically the leading open model in terms of general intelligence. Its profile is more specialised: strong in agents, cybersecurity, document processing, professional workflows and certain multimodal tasks, while other systems continue to lead in some areas of general reasoning or coding.
Mistral Is Building More Than a Model – It Is Building Its Own Infrastructure
The launch also needs to be understood in the context of Mistral’s broader development.
In September, the company raised €3 billion in a new funding round and said a significant proportion of the capital would be invested in its own European data centres, model training and product development. According to Mistral, Large 4 is the first major milestone of this new expansion phase.
Training is not necessarily finished either. Mistral says reinforcement learning is continuing and has not yet shown a clear sign of saturation. The preview version could therefore change further before the final weight release and beyond.
That is another reason to interpret the current benchmarks cautiously. Large 4 remains explicitly a public preview, and comprehensive independent testing of the final open-weight version is still largely outstanding.
Potentially Mistral’s Most Important Release Yet
Mistral Large 4 is ultimately interesting for reasons that go far beyond the headline figure of one trillion parameters.
What matters is the combination: a model trained in Europe, multimodal, agentic and multilingual, with a one-million-token context window, strong cybersecurity capabilities and the prospect of model weights that organisations can run themselves.
Mistral is therefore targeting a gap in the AI market that is becoming increasingly important. Businesses want highly capable models, but they also want greater control over their data, infrastructure and technological dependencies.
Whether Large 4 truly becomes a European counterweight to the leading American and Chinese models cannot be judged reliably just days after launch. The early results point to a capable and unusually versatile model, but not an unequivocal new global leader.
The more important moment will come at the end of October. Once the weights are released and independent developers and organisations can test Large 4 on their own infrastructure, it will become much clearer whether “Le Chonk” is more than an impressive European challenger – and whether it can genuinely become an open platform for the next generation of professional AI systems.
