Mistral Releases Open Weights for Trillion-Parameter ML4 Coding Model
French AI lab Mistral has released open weights for Mistral Large 4, a 1 trillion-parameter model nicknamed 'Le Chonk'. The model, optimized via reinforcement learning, now leads open-source coding benchmarks, surpassing China's GLM-5.3 on Terminal-Bench 4.0. It also demonstrates strong cybersecurity capabilities, matching top proprietary models while using significantly fewer tokens, marking a major step for Western open AI.
Key points
- Mistral AI, a Paris-based artificial intelligence company, has released the open weights for its Mistral Large 4 (ML4) model, previously previewed in October.
- The 1 trillion-parameter model, which activates 52 billion parameters per task, has improved its Terminal-Bench 4.0 coding score from 28% to 41% after three weeks of reinforcement learning.
- ML4 now ranks alongside or ahead of competitors like China's GLM-5.3 and OpenAI's Codex GPT-6.1 Terra on coding indices while using 47% fewer tokens on average.
- In cybersecurity tests, ML4 scored 79% on CyberGym-E2E, matching MiMo-V2.6-Pro and outperforming DeepSeek V4.1 Flash, demonstrating proficiency in patching memory-safety bugs.
- A quantized version of ML4 is available for deployment on four Nvidia B200 or B300 GPUs, aiming to provide economical alternatives to expensive proprietary AI systems.
Mistral Unveils Open Weights for Le Chonk
French artificial intelligence laboratory Mistral AI has officially released the open weights for its Mistral Large 4 (ML4) model, a development that signals a significant shift in the competitive landscape of open-source large language models. Previously introduced in a preview version on October 6, the model has undergone substantial refinement over the past three weeks. The release, now available on Hugging Face, includes the full model weights alongside new benchmark scores that highlight its enhanced capabilities in agentic coding and cybersecurity defense.
Performance Gains and Benchmark Rankings
The core of the ML4 update lies in its improved performance metrics, achieved through three additional weeks of reinforcement learning guided by customer feedback. The model, internally nicknamed "Le Chonk" due to its massive size, contains one trillion parameters but activates only 52 billion for each specific task. This architecture allows it to process text and images while producing code and text outputs efficiently.
In the realm of agentic coding, ML4 has demonstrated marked improvement. On the Terminal-Bench 4.0, which tests a model's ability to perform coding tasks within a terminal environment, ML4's score jumped from 28% in its preview phase to 41% in its final release. This performance places it slightly ahead of China's GLM-5.3, which scored 40% on the same benchmark. On the broader Artificial Analysis Coding Agent Index, which aggregates scores across multiple benchmarks, ML4 ties with GLM-5.3 and OpenAI's Codex GPT-6.1 Terra at an overall score of 48, trailing GLM-5.3's composite score of 49 by a single point.
Efficiency and Cybersecurity Strengths
Beyond raw scoring, Mistral emphasizes the model's efficiency. The company reports that ML4 rivals top open models like GLM-5.3 while using 47% fewer tokens per task on average. This reduction in token usage is particularly significant for coding agents that run for extended periods, as it lowers computational costs and latency. For enterprises relying on proprietary AI, which can be expensive to operate, economical models that maintain high performance are increasingly attractive.
The model also shows robust capabilities in cybersecurity. On the CyberGym-E2E test, which evaluates a model's ability to find, reproduce, and patch real memory-safety bugs in open-source C and C++ projects, ML4 achieved a score of 79%. This result matches the performance of MiMo-V2.6-Pro and surpasses DeepSeek V4.1 Flash, which scored 76%. These results suggest that open models are rapidly closing the gap with proprietary systems in specialized technical domains.
Deployment and Global Implications
To facilitate broader adoption, Mistral is also releasing a quantized version of ML4. This optimized variant is designed to run on four Nvidia B200 or B300 GPUs, making high-performance AI more accessible to organizations with limited infrastructure. The release underscores the growing competitiveness of Western open models against their Chinese counterparts. While benchmark scores should be interpreted with caution due to potential gaming of tests, the consistent improvements in ML4 highlight the rapid pace of innovation in the open-source AI sector.
As the AI industry continues to grapple with the high costs of proprietary models, the availability of efficient, high-performing open alternatives like ML4 may reshape how organizations approach AI integration. The model's ability to deliver strong coding and cybersecurity results at a lower token cost positions it as a serious contender in the global AI market.
Key takeaways
- Mistral AI has released open weights for its 1 trillion-parameter ML4 model, nicknamed "Le Chonk."
- The model's coding score on Terminal-Bench 4.0 improved from 28% to 41% after reinforcement learning.
- ML4 uses 47% fewer tokens than comparable models while matching their performance on coding indices.
- In cybersecurity tests, ML4 scored 79%, outperforming DeepSeek V4.1 Flash and matching MiMo-V2.6-Pro.
- A quantized version is available for deployment on four Nvidia B200 or B300 GPUs.
Sources
The WireByte editorial team synthesises technology news from multiple primary sources, verifies the facts, and links every source. Articles are produced with AI assistance and reviewed under our editorial policy.
