Key Highlights

  • Mistral Large 4 has 1 trillion parameters and launches in preview to developers, cybersecurity leaders and state authorities.
  • Mistral called it the strongest open-weight model outside China by a substantial margin, but it still lags in coding.
  • The model was trained over two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers.
  • Mistral raised 3 billion euros in a September Series D at a valuation of 21 billion euros.

A Trillion-Parameter Model for the Open Field

Mistral on Tuesday unveiled a new model it said ranks among the best open models in the world, as Western companies race to close the capability gap with Chinese developers. Mistral Large 4, also known as ML4 or "le Chonk," has 1 trillion parameters and, the French company said, is particularly effective at tasks related to cyber, coding, manufacturing, finance and multimodal work.

The model is launching in preview to developers and "cybersecurity leaders," alongside state authorities, before a wider release later this month. Mistral said that when its core parameters are released, ML4 will rank among the top open-weight models globally on aggregate benchmark performance, and called it the strongest open-weight model developed outside China by a substantial margin. The company acknowledged that the model still lags the frontier in areas such as coding.

Trained in Europe on Nvidia Hardware

ML4 was trained over two months on 4,000 Nvidia Grace Blackwell GPUs deployed in Mistral's own data centers in Europe. The company's co-founder and chief scientist said capabilities will improve as Mistral scales up its training capacity following its Series D fundraise. In September Mistral raised 3 billion euros ($3.4 billion) at a valuation of 21 billion euros, a round that cemented its status as one of the best-funded and most valuable AI startups in the region. The company has long been viewed as Europe's main answer to OpenAI and Anthropic.

The chief scientist also pointed to cyber defense, saying the model's capabilities will let enterprises and governments defend themselves against threat actors that jailbreak closed models to carry out cyber attacks.

Why Open Models Matter

Open models can be modified and self-hosted, in contrast to the leading closed systems produced by OpenAI and Anthropic. The most capable open models so far have come from China, and Chinese open models are increasingly adopted around the world. Access to them has become a flashpoint in the intensifying race for technological supremacy between the US and China, and Western developers are racing to compete in open-weight and open-source AI.

A Breach Defense Put Them in the Spotlight

Open models were thrust into view earlier this year when startup Hugging Face turned to China-based Z.ai's GLM 5.2 to defend against a breach by rogue OpenAI agents. Hugging Face first tried leading US closed systems, but their guardrails mistook its defensive actions for an attack and blocked it.

Other Western Entrants

On Monday, Nvidia-backed US startup Reflection AI unveiled its first open model, which it said matched the capability of GLM 5.2. That Chinese model was itself surpassed in August by GLM 5.3.