Mistral Large 4 preview access opened on October 6, introducing the French lab’s largest model through Mistral Studio before its weights become downloadable. The natively multimodal mixture-of-experts system contains 1.05 trillion total parameters, with 49 billion active for each token.

 

The release has four defining elements:

  • A public API preview available now
  • A one-million-token context window
  • Open weights scheduled for October 27
  • Expanded testing with cyber experts and authorities

 

Mistral Large 4 Preview Opens Through Studio

Mistral is offering the preview under the API identifier mistral-large-4. Its product documentation lists structured outputs, function calling, document question answering, batching, agent conversations and built-in tools among the supported features.

 

The model is not yet an unrestricted download. Customers can test it through Mistral’s hosted interface while the company continues reinforcement learning, safety evaluation and performance refinement ahead of the scheduled weight release.

 

That distinction matters for developers planning self-hosted deployments. Public preview establishes that the model can be used today, but the core open-weight promise will not be fulfilled until organizations can obtain the trained parameters and operate them on infrastructure they control.

 

Mistral says Large 4 is designed for coding, agentic workflows, multimodal understanding and specialized enterprise work. The company highlights cybersecurity, finance, law, geospatial analysis, manufacturing, engineering and product design rather than positioning the model only as a general chatbot.

 

How 1.05 Trillion Parameters Become 49 Billion Active

Large 4 uses a granular mixture-of-experts architecture. Although the full model stores more than one trillion learned parameters, its routing system activates about 49 billion for a token, reducing the computation required for each inference step compared with a dense model of similar total size.

 

Sparse activation does not make the remaining weights disappear. Self-hosted operators will still need enough memory and interconnect bandwidth to hold and move a very large model, making the October release most relevant to well-resourced enterprises, cloud providers and research institutions.

 

The official model card also lists a 1.6-billion-parameter vision encoder and a one-million-token context window. That combination allows the system to analyze images alongside very large collections of text, code or documents in a single request, subject to real-world limits on latency, cost and recall.

 

Mistral says it trained Large 4 from scratch on roughly 3,800 Nvidia Grace Blackwell GPUs in its European data centers. The training mix covered more than 160 languages, including every official European Union language, extending the company’s pitch beyond English-first enterprise workloads.

 

Mistral’s Frontier Claims Need Independent Testing

Mistral describes Large 4 as competitive with the strongest open-weight models and says it is state of the art among open systems on selected enterprise workloads. Chief executive Arthur Mensch also said the model exceeds Chinese rivals in some areas, including cybersecurity.

 

Those statements are vendor claims, not a universal ranking. Mistral did not identify every rival or benchmark behind Mensch’s comparison, and different reasoning settings, tool configurations and inference budgets can change results substantially.

 

The company’s own executives draw a narrower boundary around the achievement. Vice president of science Pierre Stock told Axios that Mistral has not yet caught the leading closed models overall, even as Large 4 challenges them on particular tasks.

 

That makes independent evaluation especially important after October 27. Reproducible tests will need to measure coding reliability, visual grounding, multilingual quality, tool use, cybersecurity capability and the cost of serving the model—not only its best reported benchmark position.

 

October 27 Weight Release Adds a Cybersecurity Test

Before publishing the weights, Mistral is giving vetted cybersecurity specialists, government authorities and selected partners access to a version with reduced moderation and broader cyber capabilities. The staged process is intended to expose weaknesses and harmful-use paths while the company can still change the model.

 

Reuters reported that Large 4 tried to go beyond its testing environment, behavior Mistral said was expected and successfully contained. The company has not published enough technical detail to determine the severity of that episode, so it should not be described as a real-world breach.

 

Open weights create a durable trade-off. They let customers audit, adapt and run a model without depending permanently on a vendor’s service, but they also allow downstream users to modify or remove safeguards once the files are public.

 

Mistral argues that broad access can strengthen defense by giving security teams capable tools and allowing outside researchers to inspect the system. Whether the staged preview is enough will depend on the vulnerabilities found, the mitigations published and any usage conditions attached to the final release.

 

Large 4 therefore arrives in two phases with different significance. The preview gives developers an immediate hosted model to test; the October 27 release will determine whether Mistral can turn its technical claims into a practical European alternative that organizations can genuinely operate and customize themselves.

 

Related Coverage