Mistral Large 4 goes open-weight: a trillion parameters and a gap that needs explaining

Mistral AI has released the weights of Mistral Large 4, nicknamed “Le Chonk”, in the mistralai/Mistral-Large-4-1T-A52B repository on Hugging Face. The official page gives no date. According to Vals AI and the trade press, the release came on October 29, 2026, just ahead of the end-of-month deadline announced on October 6, when the model was unveiled as an API-only preview. The model card describes it as a natively multimodal Mixture-of-Experts: it takes text and images as input and produces text. It has about one trillion total parameters and 128 routed experts, 4 of them active per token, plus one shared expert. There are 49 billion active parameters, or 52 billion if embeddings and output layers are counted, which is where the discrepancy between sources comes from. Rounding out the picture are a one-million-token context window and a built-in draft head for speculative decoding, DSpark, which Mistral says delivers up to 1.9 times the throughput, depending on hardware and load.

Compared with the preview, Mistral reports three additional weeks of reinforcement learning on coding and defensive cybersecurity, plus three reasoning modes: Low, Medium and High, with High as the default. Here are the company's own figures. On Terminal-Bench 4.0 the score rises from the preview's 28%, also self-reported, to 41%, with 30% fewer output tokens. On the Artificial Analysis Coding Agent Index the model scores 48%, level with GLM-5.3 but using 47% fewer tokens per task. Then come 19.7% on Cognition's FrontierCode 2.0 and 84.5% on BrowseComp. On safety, Mistral reports 97.1% attack resistance on B3 Agent Security and an average refusal rate of 96.6% for malicious cyber requests: these are results measured on those tests, not guarantees about how a system behaves in production. The model card credits the model with 26 points on the Artificial Analysis Intelligence Index v5, which Mistral describes as the highest score among Western open-weight models. Vals AI ranks the model 17th out of 32 on its Vals Index, at 56.53%. On the only benchmark both have measured, Terminal-Bench 4.0, it scores 26.26% (26th out of 59), against the 41% claimed by Mistral. The full technical report has been announced as forthcoming but has not yet been published.

The license, the Mistral Ecosystem Open-Weight License, is described as a modified version of MIT. It allows use, modification, fine-tuning and commercial deployment, but anyone offering the model as a service must obtain a separate commercial license from Mistral. What this clause means in practice for third-party cloud providers is not yet clear. According to the model card, training used about 3,800 NVIDIA Grace Blackwell GPUs, on European infrastructure according to Mistral. The deployment examples use an 8-GPU node with vLLM or SGLang, but these are reference configurations: no official minimum hardware requirements have been published. There is also an NVFP4-quantized checkpoint for Blackwell GPUs. Via the API, Mistral lists a price of $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. The model is also available on Mistral Studio, OpenRouter, NVIDIA endpoints and Microsoft Foundry, and Mistral says it is fluent in all the official languages of the European Union.

The release lands in an open-weight market dominated by Chinese labs, with models such as GLM, Kimi, DeepSeek and Qwen, and Mistral is pitching it as a tool for European technological sovereignty, particularly in cybersecurity. The company is announcing a specialized variant in preview for selected partners, including Orange Cyberdefense, with vetting and contractual safeguards. It adds that it will update its guardrails after observing in real traffic, in its own words, “a potential to help in conducting real attacks”. Orange Cyberdefense CEO Hugues Foulon said: “With Mistral AI, we are building European cyber AI, trained by our experts and operated under their control.” Hugging Face VP of Product Jeff Boudier commented: “AI sovereignty isn't about choosing which API to call.” A trillion open-weight parameters from a European lab is a fact. The 41% on Terminal-Bench, for now, remains Mistral's claim, and the only external measurement available says 26.26%. Before giving more weight to either number, I'm waiting for the technical report. — Olya