Mistral published open weights for ML4, a 1 trillion parameter model

What Mistral put out

The weights are the parameters a model learns during training, and publishing them lets users run and adapt ML4 themselves, heise online reported. Mistral says ML4 is its largest and most advanced model and is meant to close the gap to open-weight leaders Qwen from Alibaba, Gemma from Google and Kimi from Moonshot AI. It holds 1 trillion parameters, of which 49 billion are active for each token. Pierre Stock, Mistral's vice president of science, said in a podcast interview that the model routes each token to different sub-networks, or experts, making it expensive to train but efficient to run.

Internally, Mistral calls the model Le Chonk, Stock said, because it is a large model.

The numbers Mistral published

Mistral published the benchmark figures itself. ML4 scores 48% on the Artificial Analysis Coding Agent Index, practically level with GLM-5.3 from Z.ai, and 79% on CyberGym-E2E, level with Xiaomi's MiMo-V2.6-Pro. Heise online reported that the figures partly support Stock's claim that ML4 is the best open-weight model for coding and cybersecurity.

Stock described cybersecurity as a deliberate bet for the lab. On agentic tests, in which models plan and carry out tasks themselves, ML4 solved just under 60% of tasks, up from about 6% for its predecessor, he said. He cautioned that every benchmark can be tricked and said models should be supervised by another capable model.

Ten megawatts and a ten-month gap

ML4 was trained in Mistral's own data centres in France using 10 MW of power, which Stock called laughably little compared with what competitors run. He said ten months separated ML4 from the previous Mistral model and attributed the gap to building those data centres and to a switch to reinforcement learning.

The final version uses 50% fewer tokens at the same level of intelligence, Stock said. Between the 6 October preview and the open-weights release, ML4 improved noticeably, he said. Quantised to FP4, it runs on four GB300 GPUs, which he called still a lot but a budget that universities and research labs can carry. Stock also said ML4 is three times smaller than Kimi K3.

Why open weights

Stock said the advantage for companies is that nobody can cut off your access. He said open weights protect business continuity, citing a case where access is suddenly blocked during an attack. Mistral installs the model on customers' own GPUs and afterwards sees neither the model nor the data, which he conceded makes debugging complicated.

Stock said Mistral plans a model family: it starts with the large model, then distills it into smaller ones, sending simple recurring tasks to small models and difficult ones to large models.