Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, a hybrid instruct-and-reasoning Mixture-of-Experts with roughly 1.05 trillion parameters and about 52 billion active per token, natively multimodal, fluent in 160-plus languages, trained from scratch on European infrastructure. API pricing lands at $0.70 input, $0.07 cached input and $2.10 output per million tokens, about half the preview rate and a 63 percent per-task cost reduction on Artificial Analysis's Intelligence Index. Mistral's own release documents report Coding Agent Index 48 effectively tied with GLM-5.3 using 47 percent fewer tokens per task, Terminal-Bench 4.0 at 41 percent on Mistral methodology and 25 percent on Artificial Analysis's independent re-run, Cyber Index 51, a 96.6 percent refusal rate on malicious-cyber prompts across JailbreakBench, StrongREJECT and AgentHarm, 97.1 percent B3 Agent Security attack resistance up from 93.3 in the preview, FinWorkBench 73.3 percent and BrowseComp 84.5 percent.
FrontierCode 2.0 at 19.7 percent is the top score Cognition has published and Surge Code blind pairwise tests place Mistral Large 4 behind only Opus 5.5. Weights ship under a modified-MIT Mistral Ecosystem Open-Weight license that charges a separate commercial fee for Model-as-a-Service providers, which is the main gate between Mistral's own hosting and a competing hosted inference business. A specialized cybersecurity variant is in preview with Orange Cyberdefense under access controls.
"Open weights mean advanced defensive AI access can't be cut off by definition."— Pierre Stock, Mistral VP Science, launch briefing
Partner stack is unusually deep for a model launch
The release ships with named defensive partners quoted alongside the weights: Booz Allen's Brad Medairy, Cisco's Amy Chang, Cloudflare's Stephanie Cohen, CrowdStrike's Bartley Richardson, Elastic's James Spiteri, ENISA's Laura Heuvinck, Fortinet's John Whittle, Orange's Hugues Foulon, Rubrik's Anneka Gupta, Zscaler's Dhawal Sharma and NVIDIA's Kari Briski. The partner stack is weighted toward enterprise security integration, which maps to the Cyber Index 51 and the Orange Cyberdefense specialized variant. Mistral CEO Arthur Mensch's framing is that the business case for an open-weight 1T model sits in the enterprise-and-defense perimeter where "advanced defensive AI access can't be cut off by definition."
"AI as a tool for empowerment, not dominance."— Arthur Mensch, Mistral CEO, Large 4 launch
VentureBeat's Carl Franzen framed Large 4 as "the top open model outside China" that still trails China overall (DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32 against Mistral Large 4 at 26 on the Artificial Analysis Intelligence Index v5.0). The 63 percent per-task cost reduction against the preview, read against the Nasdaq's worst month since March 2025 and the OpenAI primary round's reset to $1.25 trillion, is the open-weight pricing anchor the hosted-API margin now has to clear.