Axios identified Mistral Large 4 as a 1-trillion-parameter multimodal model with 49 billion active parameters. Mistral says it will release the weights by the end of October.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [Mistral Newsroom](https://mistral.ai/news/mistral-large-4/)
Why it matters #
Mistral is building its open-weight pitch around a model it says can support cybersecurity work under customers' own policies. That control also makes safeguards easier to remove after weights are released; independent testing and real deployments will show what the trade-off means in practice.
Arthur Mensch, Mistral's co-founder and CEO, said Mistral would unveil a new AI model later on October 6th, Reuters reported. Axios later identified it as Mistral Large 4 and reported that Mistral had made it available through a moderated API. Mistral says it will release the model weights by the end of October.
Mistral describes Large 4 as a 1-trillion-parameter, natively multimodal model with 49 billion active parameters. In its announcement, Mistral reports 82% on one Artificial Analysis Cyber Index test and 93% of challenges solved on Cybench, a set of 40 security exercises. Those are company-reported results; independent researchers will have to wait for the weights and further technical details to test the model directly. Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who receive the same model with reduced moderation and expanded cyber capabilities.
An earlier RuntimeWire report covered Mensch's claim that the forthcoming model would challenge Chinese rivals, before Mistral identified it as Large 4. Mistral had published numerical model specifications and benchmark results before: its February 2024 Mistral Large announcement included both.
A release staged around security testing
Mistral is making the model available first through a moderated API. As it prepares the weights, Mistral says it is red-teaming Large 4 in real-world settings with cybersecurity leaders, vetted partners and state authorities. Those groups receive the same model with reduced moderation and expanded cyber capabilities.
Mistral says open weights and self-deployment can let security teams work under their own policies, including on vulnerability research and incident response that provider safeguards may block. The same access can make safeguards easier to remove once users can modify the weights. Mistral has not yet published the weights or its post-training methodology, leaving external researchers without the means to test the model directly.
Mensch and co-founders Guillaume Lample and Timothee Lacroix founded Mistral in April 2023. Mistral's account of its founding traces the idea to 2022, when the founders saw major technology companies advancing AI while closing access to their systems. Mensch worked at Google DeepMind; Lample and Lacroix worked at Meta and co-authored the LLaMA research paper. In a September interview with Le Monde, Mensch said, "AI is software. It can be controlled." Large 4's forthcoming weight release will test how much control Mistral is willing to hand to users.
Axios reported that Large 4 was trained on 4,000 NVIDIA Grace Blackwell GPUs over two months in Mistral's own European data centers; Mistral's announcement says it used 3,800 GPUs. Mistral says a significant share of its training data covered more than 160 languages. The model also uses the training, customization and reinforcement-learning environment Mistral offers customers through Mistral Forge. Mistral is selling the release as both a model and a demonstration of the tools it wants enterprise and government customers to use to customize and deploy AI.
Mistral raised €3 billion in a Series D on September 8th at a post-money valuation above €21 billion, according to TechCrunch. Large 4 puts that financing behind a visible bet on frontier models alongside Mistral's infrastructure and enterprise business.
A new Western open-weight competitor arrived the day before Mistral's model unveiling: Reflection introduced its Beam model on October 5th. Axios described both companies' models as alternatives for businesses and governments weighing closed AI systems and Chinese models. For Mistral, the weights, external testing and customer deployments will show whether its promise of user control holds up in practice.