Mistral ML4 Releases Weights and Advances in Coding

Mistral has released the weights of ML4, its model "Le Chonk" with one trillion parameters. The initial results published position it among the best open-source models in coding and show ambitions in cybersecurity, backed by training conducted in Europe with 4,000 GPUs.

ML4 trained on 4,000 GPUs in Europe, far from the 100,000 mentioned for GPT-6

ML4 was designed from scratch using 4,000 Nvidia Grace Blackwell GPUs over a two-month period within Mistral's European data centers. Developing a high-performing model in a context of limited computing resources poses a challenge for a lab the size of Mistral, especially compared to players like Anthropic. For comparison, Jensen Huang stated on X that OpenAI trained GPT-6 Astra on approximately 100,000 GPUs.

In coding, 41% on Terminal-Bench and a total of 48 on the Global Index

On the agentic coding benchmark Terminal-Bench 4.0, ML4 achieves 41%, up from 28% in its preview version, surpassing GLM-5.3 from Opencode by one point. On the Global Index, which aggregates three benchmarks, ML4 scores 48, one point behind GLM-5.3 at 49, while Codex GPT-6.1 Terra also shows 48. This progress is notable after just a few weeks of additional training. However, it is advisable to interpret these scores with caution, as some models may cheat on benchmarks.

Cybersecurity: less restricted access for testing and score above DeepSeek V4.1 Flash

Mistral highlights the cybersecurity capabilities of ML4, with customizable control and data sovereignty. In recent weeks, the company has conducted assessments of these features in cooperation with developers and officials, a practice also followed by OpenAI, Google, and Anthropic. The first groups of testers had access to a less restricted version of ML4, incorporating advanced cybersecurity features. On the CyberGym benchmark, which evaluates the detection and correction of memory vulnerabilities in C/C++ without failing the tests, "Le Chonk" slightly outperforms DeepSeek V4.1 Flash. Mistral claims to have designed ML4 so that security remains under user control. During the briefing on October 6, Guillaume Lample stated that these capabilities should enable businesses and governments to defend against malicious actors circumventing closed models.

Open alliance and risk perception: open models re-evaluated

Following the breach at Hugging Face, Mistral joined the Open Secure AI Alliance initiated by Nvidia. This collaboration across different sectors asserts that open-source models play a key role in democratizing defense mechanisms in light of the rising incidents of AI-related security breaches. While recent events have pitted open models against proprietary solutions, the former, initially perceived as less reliable, are now being considered as an alternative for protection, after closed models from Anthropic and OpenAI demonstrated similar levels of risk.

Token efficiency and usage cost: Mistral emphasizes economy

Proprietary models remain costly to use, while economical alternatives attract attention when they do not sacrifice performance. According to Mistral, ML4 competes with the best open-source models on the Artificial Intelligence Coding Agents Index, while using an average of 47% fewer tokens per task.

Release of weights, preview date, and nickname of one trillion parameters

Mistral released the weights of ML4 after launching a preview version on October 6. The model, nicknamed "Le Chonk," has one trillion parameters. The newly published benchmark scores indicate that Western open-source models are closing the gap with their Chinese counterparts.