On October 8th, French President Emmanuel Macron delivered a speech at the ‘Bpifrance Inno Generation’ event held in Paris.
Macron said: "We are engaged in the competition of cutting-edge artificial intelligence models. Mistral, a French AI company, has just released its latest model, and it has also joined this competition."
Macron said: "Just a few months ago, I still heard many people having doubts about it. Now they are already competing on the field. Which other countries in the world are involved in this battle for cutting-edge models? The United States and China, and between these two powers, there's also Canada."
Previous reports:
On October 6, 2026, local time, Mistral released the public preview version of Large 4 (ML4). The model has a total of 1 trillion parameters, with 4.9 billion parameters activated per Token. It is a native multi-modal model. The API has been made available, but the model weights will not be released until the end of this month. Prior to that, the model will undergo red-team testing in real environments with governments and cybersecurity agencies.
Mistral uses 3800 NVIDIA GB300 GPUs in its European data center to train ML4. During the subsequent training phase, Mistral revealed that it currently uses approximately 3000 GPUs for single-reinforcement learning training, generating around 33 billion Tokens per day. After filtering, approximately 16 billion of these Tokens are available for training; tens of thousands of Rollouts are generated concurrently. This round of reinforcement learning is still ongoing.
Mistral claims that its encoding, agent-based workflow, and multimodal understanding capabilities are comparable to the world’s most powerful open-source models, and significantly surpass any open-source weight models developed in the United States or Europe. In key enterprise workloads such as cybersecurity, finance, and law, Mistral is a leader among open-source models. Artificial Analysis states that its performance is on par with DeepSeek V4.1 Flash (max).
