Mistral releases weights of its 1 trillion parameter Mistral Large 4
Mistral says the model scored 48% on the Artificial Analysis Coding Agent Index, matching GLM-5.3 while using 47% fewer tokens per task on average.
Rhea Lease · AI writer ·
4 primary · 2 independent · 2 commentary · 4 vendorA giant model's weights are split open for anyone to carry away, each copy light enough to take home.Image courtesy of Fay Tograph · AI photographer, made with Grok image generation
Mistral said it is releasing the weights of Mistral Large 4 (Le Chonk), its natively multimodal model with 1 trillion parameters.[1][2]
The release follows additional reinforcement learning after a public preview, during which Mistral said gains continued to accelerate with no signs of saturation.[1]
On the Artificial Analysis Coding Agent Index, it achieved 48%, matching GLM-5.3 while using 47% fewer tokens per task on average, Mistral said.[1][3]
The model scored 79% on CyberGym-E2E, a test of whether a model can find, reproduce, and patch memory-safety bugs in C and C++ projects, according to TNW, citing Mistral data.[4]Mistral also reported it scored 82% on the Artificial Analysis Cyber Index, where several closed frontier models scored near zero because they refuse to attempt reproduction of vulnerabilities.[5]
What people on X are saying, summarized by Grok
Reaction on X splits between praise for token-efficient coding that ties a leading Chinese model and doubt this is a practical open release. Commenters say it still trails China's best open models, loses some coding tests to a closed frontier model, is far too big for a normal machine, and carries attack-misuse risk. The license is the sore point: self-hosting is fine, but selling it as a service needs a separate deal, and researchers are already dissecting how the model is built.
Posts chosen and quoted by Grok
Mistral just released the weights of a 1 trillion parameter model.
→ 52B active, 1M context
→ ties GLM-5.3 at coding agents with 47% fewer tokens
→ trails China's best open models overall, 26 vs 32
For agents, tokens per task matter more than leaderboard rank.
For me the license is the real headline of Mistral Large 4, because the new Mistral Ecosystem Open-Weight License is basically open core for model weights: you can download, modify and run it internally, but anyone offering it to third parties via API or hosted endpoints needs a separate commercial license with @MistralAI. With a 1T model almost nobody can run at home, those hosters are exactly …
Written by Grok from posts on X, with post text as Grok returned it. Not reported or checked by us, and not checked against X. The story above does not rely on it.