Mistral releases open-weight ML4 for coding & cyber

Mistral has released the weights for its open model, Mistral Large 4, on Hugging Face, describing it as its flagship open frontier system.

Also referred to as ML4, the model is a 1 trillion-parameter natively multimodal system with 52 billion active parameters. Mistral said it continued training the model for three weeks using reinforcement learning after an earlier preview, while testing it with customers and cybersecurity partners ahead of the wider release.

Mistral said the model ranks as the top open model for agentic coding on Artificial Analysis' Coding Agent Index. It also said ML4 improved its score on Terminal-Bench 4.0 from 28% to 41%, putting it level with GLM-5.3 on that benchmark.

On the Artificial Analysis benchmark, ML4 uses 47% fewer tokens per task on average than GLM-5.3 while delivering similar results, Mistral said. The model also includes three reasoning modes - low, medium and high - allowing users to set the level of reasoning for each task.

Coding Focus

Mistral is positioning the release around coding and cybersecurity use cases, particularly for organisations that want to run models on their own systems. It said the approach could appeal to regulated sectors seeking tighter control over code, data and access.

The model can also be deployed on a customer's own infrastructure for cyber defence work, according to Mistral. That would allow enterprises and government bodies to detect, analyse and respond to threats across software code and IT systems without relying on a third party to maintain access.

ML4 was trained on 3,800 Nvidia Grace Blackwell graphics processing units in Mistral's European data centres, according to the company. It said this gave it a path to state-of-the-art results with lower compute and capital demands than closed-model developers.

Mistral said one technical change since the preview was the addition of two reasoning modes during the same reinforcement learning run.

"ML4 OSS showcases our science team's ability to innovate at speed. Building on the ML4 Preview, the open weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation. In under three weeks within the same reinforcement learning run, we integrated customer feedback by introducing two additional reasoning modes, while making ML4 more efficient and customisable," said Pierre Stock, Vice President of Science, Mistral.

Open Model Push

The release comes as debate continues over whether leading AI systems should be distributed as open-weight models or kept behind proprietary interfaces. Open-weight releases allow developers and organisations to download model parameters and run them on their own infrastructure, although deploying them at scale still requires substantial computing resources.

Mistral has built its profile around open models as it competes with larger US AI groups. In comments on the release, it contrasted its funding with the spending levels associated with some rivals developing closed models.

"Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors. Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion. We have taken the lead globally in compute efficiency and have widened the gap between Mistral and other labs in that area. Our approach is to deliver quality science and put the results in everyone's hands," said Arthur Mensch, Chief Executive Officer and Co-Founder, Mistral.

Mistral also outlined several industrial uses for the model. It said manufacturing groups could use it to turn two-dimensional blueprints into three-dimensional models inside computer-aided design software, while financial institutions could apply it to spreadsheet work, large-scale data analysis and information extraction from filings.

Its multimodal design could also support geospatial tasks such as analysing satellite and aerial imagery for wildfire detection, crop monitoring and checks on power line infrastructure risks, Mistral said.

ML4 is also available through Mistral's application programming interface alongside the Hugging Face release, according to the company. The launch follows the recent introduction of Voxtral 3, a set of voice models and voice agents, as Mistral broadens its model lineup across different formats.