Mistral Releases 1T Large 4 Open Weights

Mistral AI on Thursday, Oct. 29, published the weights of Mistral Large 4, its 1-trillion-parameter flagship, on Hugging Face, making good on a promise to release the model by the end of October. The French company is pitching the open-weight model, nicknamed "Le Chonk," as a cyber-defense tool that enterprises and governments can run on their own hardware.

Key takeaways

• The weights are out under the Mistral Ecosystem Open-Weight License, a modified MIT license that is free for commercial use except for companies selling the model as a hosted service.
• The model has about 1 trillion total parameters but activates roughly 49 billion per token, with a 1-million-token context window and image input.
• Mistral reports a 96.6% average refusal rate on malicious cyber requests, yet says it still sees "a potential for uplift" for real attacks and is adjusting its guardrails.

What Mistral Large 4 open weights include

According to the Hugging Face model card, Large 4 is a mixture-of-experts model with 128 routed experts, four of which fire per token, plus one shared expert. That works out to about 49 billion active parameters, or 52 billion counting embedding and output layers, which explains the "A52B" in the repository name.

The model reads text and images and writes text, supports more than 160 languages and offers three reasoning-effort settings: low, medium and high, the default. Mistral ships FP8 weights, an NVFP4 checkpoint for Nvidia Blackwell GPUs and GGUF files for llama.cpp. Its example serving setups run on a single eight-GPU node with vLLM or SGLang.

A bundled draft head called DSpark, used for speculative decoding, can deliver up to 1.9 times the throughput of the base model alone, depending on hardware and workload, the model card says.

Mistral Large 4 benchmarks and pricing

In its release post, Mistral reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with China's GLM-5.3 while using 47% fewer tokens per task on average. Its Terminal-Bench 4.0 score rose to 41% from 28% in the preview, with 30% fewer output tokens.

The model card adds a score of 26 on the Artificial Analysis Intelligence Index v5, which Mistral calls the highest among Western open-weight models. On Mistral's API, the model costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. It is also offered on OpenRouter, Nvidia NIM and Microsoft Foundry, which Mistral says is the first cloud provider to carry it.

Why the cybersecurity pitch matters

Mistral is betting that security teams want strong models they control. "Access that can't be cut off by definition because the model is open-weighted," Pierre Stock, Mistral's vice president of science, told TechCrunch. CEO Arthur Mensch said AI should "be a tool for empowerment, not dominance."

The trade-off is that open weights cannot be recalled. Mistral says it saw the model refuse black-box attacks on live targets after the preview, but still sees potential uplift for real attackers. A specialized cyber variant remains in a vetted preview for design partners such as Orange Cyberdefense. "We don't want to fearmonger people. We think it's a problem that is solvable," Stock told TechCrunch.

What's next

The release follows the €3 billion round Mistral closed in September at a €21 billion valuation, and the model was trained on 3,800 Nvidia Grace Blackwell GPUs. With weights now public, independent labs can test Mistral's refusal and benchmark claims directly, and those results will decide whether the model earns a place in security operations centers.

FAQ

Is Mistral Large 4 free to use commercially?
Mostly. The modified MIT license allows use, modification, fine-tuning and commercialization, but businesses running a Model-as-a-Service offering need a separate commercial license from Mistral.

What hardware does it need?
Mistral's example configurations use eight GPUs on one node. An NVFP4 checkpoint targets Blackwell GPUs, and GGUF quantizations are available for llama.cpp.

How does it compare with GLM-5.3?
Mistral says the two are effectively tied on the Artificial Analysis Coding Agent Index at about 48%, with Large 4 using 47% fewer tokens per task.

References

• Mistral Large 4 open weights — Mistral AI (Oct. 29, 2026)
• mistralai/Mistral-Large-4-1T-A52B — Hugging Face model card
• Mistral releases weights for its trillion-parameter AI model targeting cybersecurity — TechCrunch (Oct. 29, 2026)