Mistral’s ‘Le Chonk’ Sets New Standard for Western Open-Weight AI in Coding and Cybersecurity

Key points

• Mistral released the weights for Mistral Large 4 ('Le Chonk'), a 1-trillion-parameter model, on October 29, 2026.
• The model improved from 28% to 41% on the Terminal-Bench 4.0 coding benchmark after three weeks of additional reinforcement learning.
• Le Chonk scored 79% on the CyberGym-E22 cybersecurity benchmark, matching MiMo-V2.6-Pro and beating DeepSeek V4.1 Flash.
• Mistral claims the model uses 47% fewer tokens per task than rival GLM-5.3, making it more efficient for long-running coding agents.
• The company argues open-weight models are crucial for Western cybersecurity, as they cannot be remotely disabled like proprietary models.
• Mistral trained the model on 3,800 to 4,000 Nvidia Grace Blackwell GPUs in Europe, using approximately 10 megawatts of power.

Background

Mistral first previewed 'Le Chonk' on October 6, 2026, positioning it as the most advanced open-weight model outside of China. This release is part of a broader Western effort to challenge Chinese dominance in open-weight AI, following the launch of Reflection AI's 'Beam' model. The timing coincides with increased concerns over AI-driven cybersecurity threats, including a summer incident where an OpenAI model breached multiple companies, and recent vulnerabilities discovered in foundational web technologies. Mistral's move aligns with a growing trend among Western labs to prioritize open-weight models for security and sovereignty, as seen in collaborations like Apple and Meta's local AI initiatives.

How outlets are covering it

WIRED emphasizes Mistral's claim that 'Le Chonk' has leapfrogged China in the open-weight race, highlighting the model's cybersecurity capabilities and the argument for open-weight models as a defense against AI-driven threats. The Next Web provides detailed technical benchmarks, noting the model's improvement from 28% to 41% on Terminal-Bench 4.0 and its efficiency gains, while also mentioning Mistral's resource constraints compared to larger labs. ZDNET focuses on the competitive landscape, noting that while 'Le Chonk' ranks highly, it is slightly behind GLM-5.3 on the overall Coding Agent Index, and highlights the broader trend of open models gaining ground against proprietary systems. All sources agree on the model's significance but differ in emphasis: WIRED on geopolitical implications, The Next Web on technical performance, and ZDNET on competitive rankings.

Why it matters

The release of 'Le Chonk' marks a pivotal moment in the global AI race, demonstrating that Western labs can produce open-weight models that rival or surpass Chinese counterparts in critical areas like coding and cybersecurity. This challenges the narrative that China dominates open-weight AI and offers Western enterprises and governments a sovereign alternative to proprietary models, which can be remotely disabled. The model's efficiency and performance gains suggest that open-weight AI can be both powerful and cost-effective, potentially reshaping the market for AI services. Additionally, the release underscores the growing importance of AI in cybersecurity, as recent breaches have highlighted the risks of relying solely on proprietary systems. Mistral's move may accelerate the adoption of open-weight models in critical infrastructure, influencing future AI policy and development in the West.

What to watch

Mistral plans to release smaller, distilled versions of 'Le Chonk' in the coming months, including one that can run on a laptop, though no specific dates have been announced. The company is also expected to publish API pricing for the model, with VP of Science Pierre Stock arguing that per-token rates may not accurately reflect the model's efficiency. Mistral aims to establish a 'cadence of the frontier' by continuously improving its open-weight models, challenging both Chinese and proprietary competitors. The broader AI industry may see increased adoption of open-weight models for cybersecurity and coding, as enterprises seek more control and reliability. Additionally, the release may prompt further discussions on AI security and the dual-use nature of open models, potentially influencing regulatory frameworks in the EU and US.