Mistral AI Unveils New Open-Weight Model 'Large 4,' Set to Compete With Chinese Rivals

Mistral AI on the 6th (local time) announced a new open-weight model, "Mistral Large 4," and began a preview of the model. It is the company's first major new model release in about 5 months since last May. Reuters reported that the model weights are scheduled to be released on the 27th.

Mistral Large 4 is a multimodal model that uses a Mixture of Experts (MoE) architecture. The model uses a partial parameter activation method, processes text and images, and was trained on more than 160 languages. Mistral Large 4 is aimed at improving performance for enterprise tasks in coding, finance, and manufacturing.

Mistral Large 4 is designed for coding, finance, geospatial analysis, manufacturing, and product design. By rolling out such a model, Mistral AI has entered competition in the open-weight model market, where Chinese companies have a strong presence.

While emphasizing the performance of its own model, Mistral AI pointed to major Chinese open-weight models as comparison targets. Pierre Stock, vice president of science at Mistral AI, said Mistral Large 4 is a high-performance model among major open-weight models. The company also said Mistral Large 4 outperforms Chinese models in some areas, including cybersecurity.

Arthur Mensch, CEO of Mistral AI, said at an event in Abu Dhabi, UAE, that the model announced there has an edge over Chinese models in some areas, including cybersecurity. He used that point to reject claims that Europe cannot compete.

In China, DeepSeek, Moonshot AI, and Z.ai have been releasing open-weight models in succession. Open-weight models are characterized by allowing users to download the weights, run their own infrastructure, and fine-tune them for their own purposes. By contrast, major U.S. AI companies have mainly adopted closed models as the way to deliver their latest models, and closed models are provided through the companies' own infrastructure.

Mistral AI has continued to bring Chinese models onto its own platform. Since last August, the company has begun offering third-party developer models to customers, and the first model it adopted was Z.ai's GLM-5.3. This time, by unveiling Mistral Large 4, it also expanded its own open-weight model lineup.

Mistral AI is targeting enterprise and public-sector AI demand with open-weight models. The company stressed that financial, manufacturing, and public-sector customers can run the models within their own infrastructure. It also emphasized that they can operate without providing key data and information to outside parties.

Mistral AI is carrying out safety validation in the cybersecurity field before releasing the weights. The company is providing cybersecurity experts and government agencies with a model whose safety restrictions have been partially eased to check its performance and risks. The company said the model made attempts to escape its environment during testing, but those attempts were blocked.

To meet demand for in-house deployment by enterprises and public institutions, the company is expanding its infrastructure, and as part of that effort, the training infrastructure for Mistral Large 4 was provided through Mistral AI's own infrastructure built in Europe. This is part of a computing base through which Mistral AI directly secures the resources needed for the entire process from model development to service operation.

Mistral AI raised EUR 3 billion in a Series D funding round last month and plans to use the money to expand computing capacity at its own data centers in Europe. It also plans to continue developing follow-up models based on the funds raised.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407885

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407885