Mistral releases the open weights of Mistral Large 4, a model with about 1 trillion parameters

Architecture and features

French AI company Mistral published the weights of Mistral Large 4 on Hugging Face on Oct. 29, after announcing a preview on Oct. 6. The model, which Mistral nicknamed “Le Chonk,” uses a mixture-of-experts (MoE) architecture with about 1 trillion parameters in total. According to the model card on Hugging Face, it has 128 routed experts and 1 shared expert, with 4 routed experts active per token. That means 49 billion parameters are active per token, or 52 billion including embeddings and output layers. The “A52B” in the repository name comes from the second figure.

According to the model card, the context length is 1 million tokens. The model accepts text and image input and produces text output, using a Pixtral-style vision encoder. Mistral describes it as a hybrid that combines instruct and reasoning modes. The reasoning effort can be set to low, medium or high, with high as the default. The company says the model is fluent in more than 160 languages.

The model was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centers. The weights are offered in FP8, along with a quantized NVFP4 version that targets Blackwell GPUs. The vLLM example on the model card uses tensor parallelism across an 8-GPU node.

Mistral’s coding claims

Mistral says Mistral Large 4 is the best open-weight model for coding. According to the company, it scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, a Chinese open-weight model, while using 47% fewer tokens per task. It scores 41% on Terminal-Bench 4.0, up 13 points from the preview. It leads Cognition’s private FrontierCode 2.0 benchmark at 19.7%. Mistral says that in Surge’s blind human preference tests it beat GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5.

All of these results are Mistral’s own measurements or results the company reports in its announcement. TechCrunch’s coverage likewise includes no independent evaluation or third-party critique.

Cybersecurity

According to Mistral’s figures, the model scores 51% on the Artificial Analysis Cyber Index, which measures finding, validating and patching vulnerabilities. The company says the model refuses an average of 96.6% of malicious cybersecurity requests, the highest rate it measures among open models. Its attack resistance on the B3 Agent Security Benchmark rose from 93.3% in the preview to 97.1%.

Mistral writes that it saw the model refuse black-box attacks on live targets but sees a potential for uplift in conducting real attacks, and that it is updating its guardrails. A specialized cybersecurity variant is in preview with design partner Orange Cyberdefense. Access requires vetting and contractual safeguards, and testing is limited to owned or authorized systems. Pierre Stock, Mistral’s vice president of science, told TechCrunch that open weights give cyber defenders access that “can’t be cut off by definition.” CEO Arthur Mensch said AI should “be a tool for empowerment, not dominance.”

License, pricing and access

The model is released under the Mistral Ecosystem Open-Weight License, a modified MIT license. It allows the model to be used, modified, fine-tuned and commercialized. Businesses that offer the model as a service (Model-as-a-Service), however, must obtain a separate commercial license from Mistral. On the Mistral API, the model ID is mistral-large-4-0, priced per million tokens at $0.70 for input, $0.07 for cached input and $2.10 for output. The model is also available through Mistral Studio, OpenRouter, Vibe Code, Nvidia-hosted endpoints and Microsoft Foundry, and the weights can be self-hosted with vLLM or SGLang.