Mistral releases the weights of Large 4 and signs Orange Cyberdefense as a design partner
Three weeks after the public preview, the Paris company published an improved version that anyone can download and run. A vetted cybersecurity variant goes first to Orange Cyberdefense.
French AI company Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, on Thursday 29 October. The 1-trillion-parameter multimodal model entered public preview earlier this month; Mistral says the released version is improved, especially in coding and defensive cybersecurity.
On the same day, Orange Cyberdefense, the security arm of Orange, said it had joined Mistral's design partner programme as the first European managed security services provider and will gradually build the model into its services.
What Mistral released
Mistral said the weights are available on Hugging Face and that the model can also be used through Mistral Studio, the Mistral API and OpenRouter. In its own launch blog, Hugging Face described a mixture-of-experts model with about 1 trillion total parameters, text and image input and a context window of 1 million tokens, shipped in FP8 and NVFP4 variants.
Reuters reported that the model can be downloaded and run independently by anyone, unlike leading systems such as Google's Gemini that are mainly accessed online. Reuters quoted Mistral's Vice President of Science Pierre Stock: "I'm not comfortable living in a world in which open weights do not exist."
Mistral said the model was trained from scratch on European infrastructure. It also introduced three reasoning modes (low, medium and high) so developers can trade cost against performance, and gave API prices of $0.70 per million input tokens and $2.10 per million output tokens.
The licence is the Mistral Ecosystem Open-Weight licence, which Mistral calls a modified MIT licence. The model is free to use, modify, fine-tune and commercialise, with one condition: businesses that run a model-as-a-service offering need a separate commercial licence from Mistral.
The cybersecurity focus
Mistral is pitching Large 4 strongly at defenders. It said the model is designed to find, reproduce and patch vulnerabilities and that its guardrails distinguish defensive from malicious requests, with an average refusal rate of 96.6% on malicious cyber prompts across three public test sets.
The company also said that while monitoring live traffic since the preview it saw "a potential for uplift to conduct real attacks" and is adjusting its guardrails. A specialised cybersecurity variant is offered only to vetted enterprise design partners, with testing limited to owned or explicitly authorised systems.
Orange Cyberdefense's role
Orange Cyberdefense said it will integrate Large 4 step by step into penetration testing, vulnerability assessment, investigation and incident response. It will rely on Mistral's computing capacity located in Europe. Reuters reported that the rollout will at first be limited to selected services for corporate and public-sector clients, that Orange Cyberdefense had €1.3 billion in revenue in 2025, and that financial terms were not disclosed.
Mistral's release also carried supportive comments from cybersecurity and technology companies including Booz Allen Hamilton, Cisco, Cloudflare, Rubrik and Zscaler, and from European users such as ABN AMRO, Belfius, La Banque Postale and ESA's Φ-lab. Microsoft said it is the first cloud provider to offer the model, through Microsoft Foundry.
Why it matters for European founders and investors
For startups, an open-weight frontier-scale model from a European company changes the build-versus-buy calculation. Teams in regulated sectors such as banking, health or public services can fine-tune and host the model themselves, which helps when customers ask where data is processed and who can switch the service off.
The licence matters for business models. Founders who embed Large 4 in their own product can use it freely, but anyone planning to resell it as a hosted model service must negotiate with Mistral. That is worth checking before pitching investors on an inference business.
Running a model of this size is not cheap. Hugging Face's own tests used multi-GPU servers, so most early-stage teams will start with the API or a cloud provider and move to self-hosting only when volumes or compliance demand it.
The same day, Germany's Schwarz Group led a Series E in Cohere, which also positions itself as a sovereign AI provider. The competition for European enterprise AI budgets is getting sharper, and cybersecurity is becoming one of the first places where open models are tested in production.
What to watch next
Mistral said a full technical report is coming. Watch for independent benchmark results, the first Orange Cyberdefense services that use the model and how Mistral updates its guardrails. The Next Web also covered the Orange partnership. For background, read our coverage of the Large 4 preview.
Spotted an error? Email contact@europreneursmedia.com and we'll correct it. See our corrections policy and editorial standards.
