Mistral Large 4 is the new flagship model from French provider Mistral AI, positioned as a European alternative to major models from the United States and China. It matters especially if you want to integrate Artificial Intelligence (AI) into a custom application through an application programming interface (API), or eventually run the model on your own infrastructure. For now, however, only a public preview through Mistral’s API is available; the promised open weights have not yet been released.
What is Mistral Large 4?
Mistral Large 4, shortened to ML4 and internally nicknamed “Le Chonk,” is a large language model (LLM) that processes and generates text. It is also multimodal, meaning that it can analyze images as well as text inputs. According to the published specifications, it has about 1.05 trillion parameters, the values learned by a model during training.
Of those, 49 billion parameters are active for each token, a small unit of text. This is possible because ML4 uses a Mixture-of-Experts (MoE) architecture, which activates only the parts of the model considered relevant to a task. That reduces computing work during use, but it does not change the need to hold the complete model in memory when hosting it locally.
According to the specifications compiled by MarkTechPost, the context window holds one million tokens. A context window is the amount of text and image material the model can consider within one request. The provider says ML4 was trained from scratch on 3,800 Nvidia Grace Blackwell graphics processors in its own European data centers, using material spanning more than 160 languages, including every official European Union language.
Mistral’s API offers two reasoning levels, “none” and “high.” These settings control how much additional processing the model applies to a complex task. According to MarkTechPost, the preview costs $1.36 per million input tokens and $4.18 per million output tokens; the supplied sources do not state subscription fees or prices specific to Switzerland.
How capable is the model?
ML4 scores 38 on the Artificial Analysis Intelligence Index. Its predecessor, Mistral Large 3, scored 9, while Mistral Medium 3.5 reached 14. That is a substantial improvement within Mistral’s lineup, although The Decoder’s assessment still places it clearly behind leading closed models such as Claude Opus 5.5.
Using the same index result, Simon Willison puts ML4 just behind DeepSeek 4.1 Flash. He considers Mistral competitive again, but still roughly six months behind the frontier. This does not directly contradict Mistral’s claim that ML4 is the best open model from Europe or the US: “best open model” and “best model overall” are different categories. The comparison also remains provisional because the weights have not been released and additional benchmark results are still pending.
Mistral emphasizes coding, cybersecurity, finance, manufacturing, and chip design. The provider reports scores of 93% on Cybench and 82% on CyberGym-E2E. Those figures have not been fully independently verified; some closed models reportedly score close to zero on security tests because their safeguards reject certain tasks, rather than necessarily because the models lack the required capability.
In practical terms, a security team could use ML4 to reproduce a known software vulnerability and then assess a proposed fix. A financial company could analyze lengthy documents together with their associated charts. Both examples reflect use cases named by Mistral, but they do not establish reliability, error rates, or suitability for production decisions.
Informal demonstrations have limitations as well. Willison asked ML4 and several competing models to create scalable vector graphics depicting an armadillo scene, using the unusual request to compare image and code output. Such tests can expose visible differences, but they cannot replace tasks drawn from your own work. An oddly dressed armadillo rarely appears in a procurement policy.
Pros and Cons of Mistral Large 4
Pros:
- European infrastructure – According to the provider, the model was trained with Mistral’s own computing resources in European data centers.
- Planned open weights – Their release could enable private installations, audits, and customization.
- Text and image support – Native image processing expands its potential use to documents, diagrams, and visual material.
- Major performance gain – The independent Intelligence Index score rises from 9 for Mistral Large 3 to 38 for ML4.
Cons:
- Not openly available yet – The model is currently accessible only through a guarded API endpoint, so self-hosting is not possible.
- Substantial hardware needs – Although only 49 billion parameters activate per token, the complete model of roughly 1.05 trillion parameters must be stored.
- Behind the overall leaders – Independent assessments continue to rank ML4 below top closed models.
- Unresolved security trade-off – Skills that help analyze vulnerabilities can support defenders, but they can also be misused.
The word “open” needs particular care. TechCrunch reports that release is planned after safety testing, around three weeks after the preview. In the meantime, Mistral intends to work with trusted partners and governments to support defensive uses while limiting malicious activity.
At launch, ML4 is therefore a planned open-weight model, meaning its model weights are intended to become downloadable, rather than a system you can already host yourself. The supplied sources also provide no license terms. Without them, it is not yet possible to determine which commercial uses, modifications, or forms of redistribution will be allowed.
How practical are the API and self-hosting options?
For beginners, the API is the realistic starting point. According to The Decoder, the preview is available to everyone through Mistral Studio, letting you test typical tasks with nonsensitive sample data: summarizing a lengthy policy, extracting information from a scanned document, or evaluating text and a chart together. You should compare quality, response time, and token use on your actual tasks rather than relying only on leaderboards.
Support in external tools has already started to appear. The llm-mistral 0.16 release adds API support for Mistral Large 4 and its reasoning mode. This is a concrete indication that outside tools are adapting to the model, although it is more relevant to advanced users and technically supported teams than to someone seeking only a finished chat interface.
Advanced users can compare both reasoning levels with a fixed set of their own evaluation tasks. Suitable examples include internal document analyses with known correct answers or image-and-text tasks whose results a subject specialist can verify. More reasoning is not automatically better: in Willison’s pelican test, the “high” result looked stronger but used only 2,717 output tokens, compared with 3,275 at the “none” setting.
Self-hosting remains theoretical for the moment. Mistral said it would release the weights by the end of October, but neither the weights nor full architecture details were available when these reports appeared. Even after release, the model’s size is likely to rule out ordinary office computers; the sources provide no precise minimum hardware specification, so credible operating costs cannot yet be calculated.
What does this mean for European and Swiss users?
ML4’s European origin distinguishes it organizationally from many large US and Chinese models. Mistral says it trained the system in European data centers using its own infrastructure. That may matter to companies and public institutions that consider a provider’s location, degree of control, and technical dependencies during procurement.
European training does not automatically mean that every API request stays inside the European Union or satisfies every Swiss data protection requirement. The supplied sources do not specify API data locations, retention periods, contract terms, or availability conditions particular to Switzerland. Those questions remain to be clarified before using confidential customer, employee, or health information.
The language coverage is promising on paper: the provider says training spanned more than 160 languages and all official EU languages. The sources do not show whether Swiss Standard German, French, and Italian perform equally well in professional settings. Schools, public agencies, and multilingual Swiss companies would therefore learn more from testing their own documents than from a general multilingual benchmark.
The financial scale also remains highly uneven. t3n values Mistral at more than €21 billion following its Series D round, while citing valuations of $852 billion for OpenAI and $965 billion for Anthropic. ML4 therefore shows that a European company can narrow the technical gap without approaching the financial scale of the largest US AI providers.
Mistral Large 4 is a credible European alternative for API trials, multilingual document work, and specialized cybersecurity or finance tasks. Its performance gain is measurable, but claims of leadership apply only to particular comparison groups and still rely partly on provider-reported figures. Its value for self-hosting will become clearer only after the weights, license terms, and safety restrictions arrive; potential misuse of its cybersecurity capabilities remains the central unresolved risk.
Sources
- Introducing Mistral Large 4: Le chonk – Simon Willison’s Weblog, 2026-10-06
- llm-mistral 0.16 – Simon Willison’s Weblog, 2026-10-06
- Mistral Large 4 – Simon Willison’s Weblog, 2026-10-06
- Mistral Large 4: Europas KI-Hoffnungsträger meldet sich mit neuem Spitzenmodell zurück – t3n, 2026-10-06
- Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model – MarkTechPost, 2026-10-06
- Mistral’s new 1T model aims to leapfrog closed and open rivals – TechCrunch, 2026-10-06
- Mistral Large 4 soll das stärkste offene KI-Modell aus Europa und den USA sein – The Decoder, 2026-10-06
Image: panumas nikhomkhai via Pexels