Mistral opens the weights of Large 4 and accelerates on code
The weights of Mistral Large 4 are now available on Hugging Face. The model gains in code and cybersecurity, but the best open Chinese models remain ahead on the overall index.
Mistral AI published on Thursday, October 29, 2026, the weights of Mistral Large 4, its 1 trillion parameter model nicknamed "Le Chonk." This release on Hugging Face allows running the model locally, without sending data to Mistral, while the startup says it has further improved its performance in code and cybersecurity after three weeks of reinforced training. The topic also matters for the open model market, where Chinese competition maintains the lead on the overall Artificial Analysis index.
A giant model, but not as demanding as it seems
On paper, Mistral Large 4 relies on 1 trillion parameters. In practice, it only uses 49 billion per token, or 52 billion if input and output layers are included. This discrepancy comes from its MoE architecture, for Mixture of Experts, a system that distributes the load among several specialized sub-models. Here, Mistral talks about 128 experts, of which only four are activated per token, plus one common expert.
The logic is not new at Mistral, but it takes on a new scale with this version. The predecessor, Mistral Large 3, already had 675 billion parameters, with 41 billion used per token. The difference is therefore less about the raw number than about efficiency in use. And that is often where the real battles lie, far from conference slogans.
What changes after three weeks of reinforcement
Between the preview presented on October 6 and the release of the weights, Mistral says it continued reinforcement learning, especially on code and cybersecurity. According to its figures, the model's score on Terminal-Bench 4.0 rose from 28% to 41%. On this command-line programming test, Mistral Large 4 only leads the Chinese GLM-5.3 by one point, with 40%, but it outperforms DeepSeek V4.1 Flash (31%), Kimi K3 (21%), and Muse Spark Open (6%).
On the Artificial Analysis code index, the model scores 48%, tied with GLM-5.3, while consuming on average 47% fewer tokens per task. Mistral also emphasizes "considerable improvements" and progress that "continues to accelerate, with no sign of saturation." The wording is ambitious, but it reflects a simple reality: the race is no longer just about size, it is about reasoning quality and cost per query.
A real opening, but with limits
The release of the weights changes the game for companies wanting to host the model on their own servers. It reduces dependence on Mistral's API and avoids outsourcing data to the publisher. At the same time, the API price is halved compared to the preview: $0.70 per million input tokens and $2.10 per output million, versus $1.36 and $4.18 previously.
However, this openness is not total. The license, called Mistral Ecosystem Open-Weight License, allows commercial use except for companies reselling access to the model as a service; they must then sign a separate license with Mistral. And, above all, the model remains out of reach for a standard PC: even its compressed version requires four Nvidia B200 or B300 GPUs, versus eight for the full version. Local, yes. On the living room machine, no.
Chinese open models maintain the overall lead
Mistral acknowledges in its documentation that the best Chinese open models remain ahead on the overall Artificial Analysis index. Le Chonk scores 26 points, the best among Western open models, but not the best in the overall ranking. This hierarchy matters, as it influenced some recent decisions, including Ecosia's, which replaced Mistral in early October with Chinese models like Qwen, Kimi, and GLM.
Christian Kroll, Ecosia's CEO, criticized Mistral for the quality of its models and a delay of about one year compared to the competition. In this context, releasing the weights is not enough to close the gap, but it puts Mistral back in the game on a key field: real-world uses, from code to private execution environments. The market will now look less at the model's name than at its ability to keep pace without exploding costs or servers.
By the numbers
- 1 trillion — parameters announced for Mistral Large 4, according to Mistral, October 29, 2026.
- 41% — Terminal-Bench 4.0 score after reinforcement, versus 28% before, according to Mistral.
- $0.70 — price per million input tokens, versus $1.36 previously.
- 4 GPU cards — minimum required for the compressed version, according to Hugging Face documentation.