Paris-based Mistral AI on October 29 released the full open weights of Mistral Large 4, nicknamed "Le Chonk" — a 1-trillion-parameter, natively multimodal mixture-of-experts model with roughly 52 billion active parameters and a 1-million-token context window. The weights landed three weeks after the model's October 6 preview, and Mistral says the reinforcement-learning run continued through that period, lifting its Terminal-Bench 4.0 score from 28.3 percent to 41 percent. The release is the clearest sign yet that Europe intends to compete at the open-weight frontier rather than cede it to American and Chinese labs.
A Trillion-Parameter Model Built for Coding and Cyber Defense
The architecture follows the preview checkpoint closely: about 1 trillion total parameters, 128 routed experts with four active per token plus one shared expert, MLA attention with a DeepSeek-style sparse-attention indexer, a 1.6-billion-parameter Pixtral vision encoder, and a context window of one million tokens. Reasoning effort now ships in three modes — low, medium and high — with high as the default. Official checkpoints are provided in FP8 and NVFP4, alongside day-one GGUF quantizations from llama.cpp and Unsloth. Mistral positions the model squarely at developers, describing it as built for coders and "trusted by defenders," with endorsements from Booz Allen, Cloudflare and Orange Cyberdefense.
Benchmark Gains and the Efficiency Argument
The headline number is Terminal-Bench 4.0, which rose from 28.3 percent at preview to 41 percent at release — a 13-percentage-point jump — while output tokens fell by 30 percent. On Cognition's private FrontierCode 2.0 benchmark, Le Chonk leads with 19.7 percent, ahead of DeepSeek V4.1 Flash at 19.3 percent, GLM-5.3 at 17.8 percent and Kimi K3 at 17.3 percent. On the Artificial Analysis Coding Agent Index it scores 48 percent, tied with GLM-5.3 while using 47 percent fewer tokens per task. It also reports 84.5 percent on BrowseComp, 73.3 percent on the finance benchmark Finch, and 97.1 percent attack resistance on the B3 agent-security benchmark.
Independent testing is more sobering. On the Artificial Analysis Intelligence Index v5, Large 4 scores 26 — the highest among Western open-weight models, but only fifth among open weights overall, with Chinese models such as DeepSeek V4.1 Flash and GLM-5.3 still ahead on the aggregate. Its 51 percent Cyber Index score ranks third among open models, behind DeepSeek and MiMo. Mistral's "best open-weight model for coding" framing holds in its own charts, but not across every independent leaderboard.
Pricing, Licensing and Deployment
API pricing is 0.70 dollars per million input tokens, 0.07 dollars per million cached input tokens and 2.10 dollars per million output tokens — roughly half the preview price. The license is a modified MIT, branded the "Mistral Ecosystem Open-Weight License": free to use, modify, fine-tune and commercialize, except that businesses offering the model as a service need a separate commercial license. On hardware, Hugging Face's tests show 2-bit quantizations fit on a single Apple Silicon Ultra machine with 512 gigabytes of unified memory, while vLLM with the NVFP4 checkpoint requires Blackwell GPUs.
What It Means
The release sharpens two debates. The first is sovereignty: Mistral's European-trained weights give European enterprises and governments a domestic option as procurement rules increasingly favor trusted, jurisdiction-bound suppliers. The second is terminology. Because model-as-a-service providers must negotiate a separate commercial license, many developers argue the result should be called "open weights" rather than "open source." Either way, a flagship that keeps improving after its announcement — with the gains landing exactly where they hit the bill, in token efficiency and deployment threshold — is a meaningful competitive signal.
Key Facts
Released October 29 2026 three weeks after the October 6 preview
1 trillion total parameters with about 52 billion active per token
1 million token context window with native multimodal input
Terminal-Bench 4.0 rose from 28.3 percent to 41 percent after three more weeks of reinforcement learning
API price 0.70 dollars per million input tokens and 2.10 dollars per million output tokens
Modified MIT license in which model-as-a-service use requires a separate commercial license
Scores 26 on the Artificial Analysis Intelligence Index v5 the top Western open-weight model but fifth among open weights overall
Sources: Mistral AI, The Decoder, AILog, Hugging Face, threerouter
Top comments (0)