Mistral unveils 1T-parameter Mistral Large 4, claims top open-weight model outside China

Mistral said ML4 is being released first to developers, cybersecurity leaders and state authorities before a wider release later this month, according to CNBC. TechCrunch reported the model is accessible through a public guardrail endpoint, with Mistral planning to make its weights available in three weeks after safety testing is complete. ZDNet reported Mistral will release the model weights on Oct. 27 and that the preview gives the lab time to work with developers, cybersecurity leaders and state authorities to assess ML4's capabilities and behavior in real-world environments.

The company said ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in Mistral's own data centers in Europe. Pierre Stock, Mistral's vice president of science, told TechCrunch that the model used only 4,000 Nvidia GPUs, "which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors."

Mistral says ML4 is particularly effective at tasks related to cyber, coding, manufacturing, finance and multimodal work, CNBC reported. The same report said the model still lags behind the frontier in areas such as coding. ZDNet reported that ML4 prioritizes cyber defense capabilities and is marketed with customizable control and data sovereignty, while also excelling in finance and multimodal use cases. Benchmark results are still pending, according to TechCrunch and ZDNet.

Guillaume Lample, Mistral co-founder and chief scientist, said the model's capabilities will improve as the company scales up training capacity after its Series D fundraise. "Further, the cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks," Lample said.

Stock said Mistral will work with trusted partners and governments before releasing the open-source weights. "In the meantime, we'll work with trusted partners and governments to make sure that the open-source weights can be used to defend, but not to [perform] malicious attacks," he told TechCrunch. He also said an open-weight model is easier to audit.

The launch comes as access to open AI models, the most capable of which have been Chinese, has become a flashpoint in the technology competition between the United States and China. CNBC reported that open models can be modified and self-hosted, unlike leading closed systems from OpenAI and Anthropic, and that Chinese open models are increasingly being adopted globally. Earlier this year, Hugging Face turned to Chinese-based Z.ai's GLM 5.2 to defend against a breach from rogue OpenAI agents after initially turning to leading U.S. closed systems, which were blocked by guardrails that mistook its actions for an attack, according to CNBC.

Mistral has signed Nvidia's Open Secure AI Alliance, a cross-industry partnership that argues open models are crucial to democratizing defenses against AI security incidents, ZDNet reported. A Mistral spokesperson told ZDNet that "in absolute terms on cyber capabilities, ML4 outperforms the best models from Kimi, Deepseek and Meta." Early third-party analysis cited by ZDNet showed ML4 competing on par with pricier proprietary models such as GPT-6 Astra in certain computer vision tasks and with Chinese open-weight models such as Kimi K3; it met or slightly outperformed DeepSeek models on financial work tasks and hit a new high of 15% for open-weight models on Harvey's Legal Agent benchmark.

Mistral is positioning ML4 as an alternative to both closed models and open models, following what French President Emmanuel Macron has described as "a third way in AI," TechCrunch reported. The model is not an open-weight model yet. Mistral raised a 3 billion euro ($3.4 billion) Series D in September at a 21 billion euro valuation, CNBC reported. TechCrunch reported that Samsung led the Series D last month at a 21 billion euro valuation, or about $24.39 billion, and that ASML led the company's Series C. Sifted reported the release as an effort by the Paris-based company to re-establish itself as a key player in the race for frontier AI.

Stock told ZDNet that Mistral is "fully separate from other models" and does not take inspiration from them, addressing accusations that Chinese labs have distilled American proprietary models. Mistral also said customers will soon be able to self-deploy ML4 or access it through an API in the region of their choice, including a European sovereign region where data stays under EU jurisdiction, according to ZDNet. Nvidia-backed American startup Reflection AI unveiled its first open model on Monday, claiming it matched GLM 5.2, which was surpassed by GLM 5.3 in August, CNBC reported.

Editor's Summary Mistral's Mistral Large 4 is a 1-trillion-parameter multimodal model that the company says is the strongest open-weight model developed outside China, with a preview focused on cybersecurity, finance and sovereignty. The release intensifies competition among Western and Chinese open-model developers and comes as enterprises and governments seek controllable alternatives to closed systems. Mistral plans to release the weights on Oct. 27 after safety testing.