Mistral launches Large 4 (Le Chonk), a 1-trillion-parameter open-weight model
Mistral's 1-trillion-parameter Le Chonk goes open-weight in three weeks.
Paris-based Mistral AI has launched a public preview of Mistral Large 4, which the company has officially nicknamed le Chonk and informally ML4. Announced on 6 October 2026, ML4 is described by Mistral as a 1 trillion-parameter, natively multimodal model with 52 billion active parameters and positions itself as a new frontier for open-weight performance.
The model is a hybrid instruct-and-reasoning Mixture-of-Experts (MoE) with multimodal input, and is designed to unify instruction-following, reasoning, and agentic capabilities in a single model. Mistral says it is state-of-the-art among open weights on tasks in cybersecurity, finance, and manufacturing, and is natively fluent in more than 160 languages.
What makes Large 4 different
ML4 is Mistral's largest and most capable model to date. Its sparse MoE design keeps the parameter count at roughly a trillion while activating only around 52 billion at a time, a structure that keeps inference costs far below what a dense 1T model would typically demand. The company framed the launch as a deliberate middle path: not the dense closed frontier models from the largest US labs, and not the smaller open models that have dominated much of the open-weight field.
Mistral's stated intent is to position ML4 as an alternative to both dominant Chinese open models and top closed competitors. Le Chonk is definitely not small, the company acknowledged, but it is not yet an open-weight model. For now it is accessible via a public preview API on Mistral Studio, with the full weight set scheduled to drop at the end of the month after additional safety testing.
Benchmarks and strategy
Mistral says the model demonstrates exceptional performance across coding, agentic workflows, and multimodal understanding, with benchmark results still maturing as training is refined. The goal, per the company, is to be best in class among open-weight models, particularly outside of China, and to leapfrog closed source competitors on focused training tasks.
- 1 trillion parameters, natively multimodal, with 52 billion active
- Hybrid instruct-and-reasoning MoE unifying instruction, reasoning, and agentic behaviour
- State-of-the-art among open weights on cybersecurity, finance, and manufacturing tasks
- Natively fluent in 160+ languages
- Public preview now on Mistral Studio; open weights due at the end of October
- Intended to serve as the foundation for a new generation of specialised Mistral models
ML4 is expected to serve as the foundation for a new generation of smaller, specialised, and cost-optimised Mistral models, a strategy the company has followed before, distilling and tuning the large model into targeted variants for production use.
If you are building on open-weight models and want something that is both large and practical, ML4's combination of a 1T parameter count, 52B active budget, and open-weight release is a notable step for the open-weights ecosystem. For enterprise teams, native multimodality and strong performance in high-stakes domains like cybersecurity and finance make it a model worth watching as the full weights arrive.
Sources: Mistral AI: Mistral Large 4 (Le Chonk); TechCrunch: Mistral's new 1T model aims to leapfrog closed and open rivals; CNBC: Mistral unveils new AI model it says rivals best open systems from China.
Corrections: Report an error at corrections@wireshaman.com.