Mistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model
Mistral’s positioning around coding and cyber-defense workloads makes those operational details more consequential. A model used for agentic coding inside enterprise environments will likely need controlled tool execution, sandboxed terminal access, policy enforcement around repositories and secrets, and auditable traces of model actions. Running the model on-premises can reduce external exposure, but it does not remove internal risk; it increases the importance of guardrails, identity integration and runtime isolation.
The reported token-efficiency gains also deserve attention beyond benchmark value. Lower tokens per task can materially affect total cost of ownership, latency and concurrency for large-scale developer-assistance deployments. But teams should evaluate whether the model’s multiple reasoning modes introduce new routing logic in applications, such as when to invoke deeper reasoning versus keeping requests fast and cheap. That is an architecture decision, not just a prompt choice.
For IT leaders, the strategic issue is whether open-weight frontier models are mature enough to justify building a durable internal AI platform around them. The upside is sovereignty and customization; the trade-off is that your organization now owns more of the performance, safety and lifecycle-management burden.
Mistral AI SAS today announced the availability of the open weights of its frontier model Mistral Large 4, nicknamed “Le Chonk,” the company’s most capable large language model to date, initially released in preview three weeks ago.
With 1 trillion parameters and 52 billion active parameters, ML4 is natively multimodal and has received three additional weeks of reinforcement training and customer testing. The company says it’s becoming one of the best open agentic coding models in the world, offering leading performance at a small fraction of the token cost with state-of-the-art cyber defense capabilities.
“ML4 OSS showcases our science team’s ability to innovate at speed,” said Vice President of Science Pierre Stock. “Building on the ML4 Preview, the open-weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation.”
The company trained ML4 on 3,800 Nvidia Corp. Grace Blackwell graphics processing units in its European data centers to achieve high-fidelity performance for tool use and coding, driven by end-to-end control of its infrastructure and the ability to optimize for hardware during model development.
Open weights mean companies can take the model and deploy it on their own hardware, within their own firewalls, and run sensitive data without exposing it to the cloud or outside parties. With Mistral’s model, companies will have a Western-developed model to compete with the multitude of ultra-large Chinese coding models on the market, such as the GLM family from Beijing Zhipu Huazhang Technology Co. Ltd. and the Kimi family from Beijing Moonshot AI Technology Co. Ltd.
“Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors,” said co-founder and Chief Executive Arthur Mensch. “Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion.”
Because of its training regimen, ML4 stands out among many other frontier models, rivaling even top models such as Chinese-built GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task on average.
On Terminal-Bench 4.0, which tests agentic coding in the terminal, ML4 climbed from 28% to 41%, tying GLM-5.3 as the best model for coding agents.
Mistral offers three reasoning modes for ML4: low, medium and high. Teams can use this to match reasoning depth to the task without adjusting prompts, which can be less effective, allowing reasoning to fit quick everyday tasks to deep work on extremely complex problems. This allows developers and enterprise users to control token use per task.
The company said the release of ML4 represents a new generation of Mistral models spanning all modalities. It follows the recent launch of Voxtral 3, a family of voice models and voice agents. Mistral plans to continue to release additional models from this family in the coming months.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.