- Published on
Mistral AI Open-Sources Large 4 Model Weights: 1T Parameters, More Token-Efficient Coding Agents
- Authors
- Name
- DeepHorus
On October 29 local time, the French artificial intelligence company Mistral AI officially opened the model weights of Mistral Large 4 (abbreviated as ML4, codenamed le Chonk). The model has 1 trillion total parameters and 49 billion activated parameters, and is a native multimodal model. Mistral said that this released version has been optimized and upgraded after real-world testing with government and enterprise users.
The research and development and training of ML4 were both completed in Mistral's own data center in Europe, using 3,800 NVIDIA Grace Blackwell GPUs to train from scratch. Compared with the initial version announced three weeks ago, this version with open weights has been optimized and upgraded in terms of performance. The model's activated parameters, including the embedding layer and output layer, are 52 billion, and the parameter scale reaches the 1 trillion level.
In the Artificial Analysis Coding Agent Index, ML4 scored 48%, basically on par with the leading open-source model GLM-5.3. The company particularly emphasized that the model consumes 47% fewer Tokens per task on average. This means that, with similar coding agent performance, ML4's invocation cost and inference overhead are expected to be significantly reduced.
For online use, ML4 is priced at $0.70 per million input Tokens, $0.07 per million cached input Tokens, and $2.10 per million output Tokens. Roughly converted at current exchange rates, these are about 4.7 yuan, 0.47 yuan, and 14 yuan RMB, respectively. This pricing system is aimed at online invocation scenarios and complements the open weights.
The model is released under a modified MIT license, allowing free use for personal, academic, modification, fine-tuning, and conventional commercial purposes. However, for cloud service providers or vendors that offer MaaS for resale, they must separately obtain a commercial license from Mistral before commercial use. This arrangement takes into account both the open ecosystem and the licensing boundary for commercial resale.
For developers, ML4's open weights and lower Token consumption provide a new choice for coding agent development. For the open-source model ecosystem, opening the weights of a 1-trillion-parameter-class model may intensify competition with leading open-source models such as GLM-5.3. Mistral, in turn, can cover different user needs through online services and commercial licensing.
In the future, ML4's actual performance in government, enterprise, and cloud service scenarios remains to be observed. Whether its modified MIT license can attract more cloud vendors to obtain commercial licenses will also affect the scope of the model's deployment. As demand for coding agents grows, Token efficiency may become one of the key indicators in model competition.
