French AI lab Mistral has released the open weights of Mistral Large 4, its flagship 1-trillion-parameter, natively multimodal model, making one of the world's most capable AI systems freely downloadable days after Google chose a tightly restricted rollout for its own frontier release. The announcement, published on Mistral's blog under the model's unofficial nickname "Le Chonk," confirmed that the weights are available on Hugging Face for self-hosted deployment with vLLM or SGLang.
The release marks a rapid transition from preview to open availability. Mistral introduced Mistral Large 4 in public preview earlier this month, and says just three weeks of additional reinforcement learning since then drove what the company describes as massive improvements in coding and defensive cybersecurity. According to Mistral, gains "continue to accelerate, with no signs of saturation," and the improved model also delivers substantial gains in reasoning efficiency that lower inference costs for developers. For more context on this story, see our ongoing artificial intelligence updates.
Benchmark Claims: The Open-Weight Coding Crown
The centerpiece of Mistral's announcement is a claimed lead on Cognition's private FrontierCode 2.0 benchmark, which evaluates end-to-end code quality and, in Mistral's words, whether a repository maintainer would merge the agent's change. Mistral Large 4 scored 19.7%, ahead of DeepSeek V4.1 Flash at 19.3%, GLM-5.3 at 17.8%, Kimi K3 at 17.3%, and GLM-5.3 Flash at 14.2%.
On the Artificial Analysis Coding Agent Index, Mistral reports a score of 48%, effectively tied with the leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average — a efficiency gap that matters greatly for anyone running coding agents at scale. VentureBeat's coverage of the release highlighted a 63% lower cost-per-task figure, and The New Stack reported roughly a one-point coding lead over GLM-5.3.
Terminal-based workflows improved as well. Mistral says the model jumped 13 percentage points on Terminal-Bench 4.0 compared with its public preview, reaching 41% and taking the lead among open-weight models while reducing errors in terminal environments.
Built for Coders, Positioned for Cyber Defenders
Cybersecurity is the second pillar of the release, and it is no accident. Reuters reported that Mistral is explicitly targeting cybersecurity users with the model, and Fortune's writeup described it as the best open model for agentic coding and cyber defense. The positioning dovetails with Google's decision this week to give its own cybersecurity-focused frontier model first to cyber defenders — but with the opposite distribution philosophy: where Google is phasing access through a trusted-partner program, Mistral is letting any organization download the weights outright.
The Sovereign AI Pitch
Mistral frames the release squarely around its "Sovereign AI" strategy. The company says the model was trained from scratch on European infrastructure with "clear lineage from inception to release," giving organizations a foundation they can own, customize, and control on their own hardware. That argument has grown louder among European enterprises and governments weighing dependence on US-based model providers.
The open-weight move also earned attention from across the ecosystem. Jeff Boudier, VP Product at Hugging Face, was quoted in the announcement saying developers need the freedom to decide where a model runs and how it is optimized, welcoming the release as a foundation organizations can build on and adapt.
ZDNET reported that the model now ranks alongside the best proprietary coding systems, while VentureBeat called it the top open model outside China — a notable framing given how dominant Chinese open-weight releases from DeepSeek, Z.ai, and Moonshot have been in recent benchmark cycles. Mistral is positioning itself as Europe's answer to both camps.
How to Get It
Developers can use Mistral Large 4 through Mistral Studio, the Mistral API, or OpenRouter. For full control, the open weights can be downloaded from Hugging Face and deployed on self-managed infrastructure with vLLM or SGLang. Mistral says the model card and license contain everything needed to get started, with a full technical report promised soon. The model can also be tried in Mistral's Vibe Code environment.
A Counterpoint to This Week's Closed Release
The timing of the release is pointed. Earlier this week Google announced Gemini 4 Carbon, a model that topped the Artificial Analysis Intelligence Index v5.0, but said access would start with a subset of cyber defenders and expand gradually through its Fairwind Program. Mistral's counter-message is straightforward: the frontier doesn't need to be locked up. With more than 1,100 upvotes on the Hacker News thread about the release within hours, the developer community's appetite for genuinely open frontier weights is clearly intact.
Whether Mistral Large 4 holds its coding lead is an open question — rival labs iterate quickly, and independent verification of vendor-reported benchmarks is always warranted. But for organizations that need strong coding and defensive-security performance under their own roof, the calculus just changed: the best open-weight coding model, by Mistral's account, is now a download away.
_This article is based on Mistral's official announcement and reporting by Reuters, VentureBeat, ZDNET, Fortune, SiliconANGLE, The New Stack, and ZDNET._
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
