Mistral Large 4 gains openness and capacity on code

Mistral released the weights of Mistral Large 4 on Hugging Face on October 29, 2026. The multimodal model advances in code and cybersecurity but remains out of reach for a standard PC.

On October 29, 2026, Mistral published the weights of Mistral Large 4 on Hugging Face, its multimodal model nicknamed “Le Chonk.” This release allows running it on one’s own infrastructure while maintaining a model presented as one of the strongest currently available for code and cybersecurity. Timing also matters for the market: three weeks after the preview, the French startup shows measurable gains and embraces a controlled open-license strategy.

Un modèle ouvert, mais pas léger

Mistral Large 4 is based on a Mixture of Experts (MoE) architecture: the model has a total of 1 trillion parameters but activates only about 49 billion per token, according to the published documentation. In practice, this distribution aims to focus computation where it is useful instead of running the entire network for each response. This helps explain why “open weights” does not mean “runs smoothly on a laptop.”

The released version comes in two official variants, FP8 and NVFP4, with GGUF conversions also available via Unsloth and the llama.cpp team. Mistral specifies that its model is distributed under the Mistral Ecosystem Open-Weight License, derived from MIT, with a restriction on model-as-a-service offerings. In other words, openness exists but comes with commercial safeguards.

Le code et la cybersécurité au cœur du positionnement

In terms of use cases, Mistral emphasizes two: software development and cybersecurity. In its communication and coverage by The New Stack, the company states that Mistral Large 4 particularly targets code agents, vulnerability analysis, and detection rule generation. The model has also been presented as useful for modernizing legacy codebases, a task that rarely makes headlines but pays the bills.

Figures provided by Mistral show clear progress since the early October preview. Terminal-Bench 4.0 rose from 28% to 41% after three weeks of reinforced training. On the Artificial Analysis code index, the model reaches 48%, tied with GLM-5.3, while consuming on average 47% fewer tokens per task. In the CyberGym E2E cybersecurity test, Mistral cites 79%, versus 29% for GLM-5.3, while the vulnerability reproduction and correction score climbs to 82% in data sourced by Hugging Face.

Pourquoi Mistral met autant l’accent sur l’ouverture

For Mistral, the challenge goes beyond performance alone. Pierre Stock, the company’s scientific vice president, defended the idea that open weights guarantee business continuity, transparency, and auditability. He also summarized the core message in one phrase: “Our stance is that AI is going to be like electricity.”

This logic resonates with teams fearing dependence on an external API at the wrong moment, especially during a security incident. With open weights, the user controls both deployment and usage policy. Mistral still refuses dangerous requests and reports a refusal rate higher than other open models on several cybersecurity test sets. The trade-off is clear: more control, but not carte blanche.

Une ambition industrielle qui dépasse le simple lancement

The weight release follows a preview phase marked by training on about 3,800 Nvidia Grace Blackwell GPUs, according to The New Stack, and rapid improvement in results. Mistral says it dedicated about 10 megawatts to this reinforced training phase, an order of magnitude it presents as lower than better-funded competitors. The company also stresses its long-term goal to “own the entire stack.”

This ambition is already reflected in the roadmap: smaller versions of Mistral Large 4 are expected in the coming months, including a variant designed to run on laptops. Meanwhile, the model is already compatible with Transformers, vLLM, llama.cpp, and Hugging Face Inference Endpoints. Companies wanting to test openness without building a complex setup thus have several options. Others will mainly have to count their GPUs.

Key points

  • Weights published on October 29, 2026 on Hugging Face.
  • 1 trillion parameters total, 49 billion active.
  • Terminal-Bench 4.0 rises from 28% to 41%.
  • GLM-5.3 remains ahead on the overall open weights index.
  • Mistral cites 3,800 Nvidia Grace Blackwell GPUs for training.
  • “Le Chonk” targets code, cyber, and multimodality.

By the numbers

  • 1 trillion parameters — total model size, according to Mistral, 2026.
  • 49 billion active parameters — activated per token, according to Mistral, 2026.
  • 41% — Terminal-Bench 4.0 score after three weeks, according to Mistral, 2026.
  • 47% fewer tokens — savings announced versus GLM-5.3, according to Mistral, 2026.
  • 3,800 GPUs — training resources cited by The New Stack, 2026.

To read