French artificial intelligence start-up Mistral AI unveiled its new flagship model, Mistral Large 4, on Tuesday, October 6, saying the release "narrows the gap" with the best openly available models in the world — a segment currently led by Chinese companies.
Co-founder Guillaume Lample presented the model as a comeback for the Paris-based company. Its last major AI release had been in December 2025, with the most recent update in April 2026, and the company's models had since slipped to 24th place in the aggregate ranking compiled by the independent reference Artificial Analysis. That ranking is led by American companies such as OpenAI, Anthropic and Google, followed by Chinese firms including Moonshot.AI, Alibaba and Z.AI.
What Mistral is claiming
According to Mistral AI, Large 4 is "among the world's best open models" — models that are freely accessible and modifiable — and is "essentially at the same level as the best Chinese models from a month or two ago."
The company released preliminary figures from thematic rankings, saying Large 4 scored 63% on Deep SWE 1.1, which measures performance on long coding tasks. That would put it on par with the much-discussed GLM 5.3 model from China's Z.AI, ranking 12th overall, against top models reaching 74%. Mistral also claimed Large 4 matches the best open models on certain specialised tasks such as finance — including handling large data spreadsheets — cybersecurity and geospatial analysis, such as identifying objects in satellite imagery to assess natural disaster damage. The company said it is also competitive in industrial design and production, a field where it aims to be aggressive, naming customers such as Airbus and BMW.
Le Monde notes these performances remain to be confirmed by the regularly updated rankings of independent third parties.
Release plan and the road ahead
The company promised detailed results on October 27, when Mistral Large 4 is due to be released as open source. Until then, security testing will be completed, the start-up said, at a time when the sector has been rattled by incidents linked to AI models. Before the open-source release, the model can be accessed through an interface that Mistral AI says it is monitoring to prevent misuse. It will eventually become the default model for the company's assistant, Vibe, though users will still be able to choose Chinese models such as GLM, which continue to outperform in some areas.
The announcement answers suspicions that Mistral was abandoning the expensive frontier-model race after its summer 2026 investments to build its own data centres in Europe. CEO Arthur Mensch told Le Monde in September that pushing model capabilities remains the company's number one pillar, and the start-up says it is optimistic about improving performance in Mistral Large 5 and 6 over the coming year on the same technical architecture.
Lample said Large 4 was trained on 4,000 state-of-the-art chips — "two to three times fewer" than what Chinese start-ups can access, and far below the hundreds of thousands used by American leaders. Mistral aims to reach data-centre capacity of 1 gigawatt by 2030, shared between training its own models and running AI for clients, and is counting on chips coming online over the next six months following its two most recent fundraising rounds, including a €3 billion raise in early September.
Sources:
- Le Monde: Mistral AI unveils new AI model aimed at 'narrowing the gap' with top Chinese competitors
Co-founder Guillaume Lample presented the model as a comeback for the Paris-based company. Its last major AI release had been in December 2025, with the most recent update in April 2026, and the company's models had since slipped to 24th place in the aggregate ranking compiled by the independent reference Artificial Analysis. That ranking is led by American companies such as OpenAI, Anthropic and Google, followed by Chinese firms including Moonshot.AI, Alibaba and Z.AI.
What Mistral is claiming
According to Mistral AI, Large 4 is "among the world's best open models" — models that are freely accessible and modifiable — and is "essentially at the same level as the best Chinese models from a month or two ago."
The company released preliminary figures from thematic rankings, saying Large 4 scored 63% on Deep SWE 1.1, which measures performance on long coding tasks. That would put it on par with the much-discussed GLM 5.3 model from China's Z.AI, ranking 12th overall, against top models reaching 74%. Mistral also claimed Large 4 matches the best open models on certain specialised tasks such as finance — including handling large data spreadsheets — cybersecurity and geospatial analysis, such as identifying objects in satellite imagery to assess natural disaster damage. The company said it is also competitive in industrial design and production, a field where it aims to be aggressive, naming customers such as Airbus and BMW.
Le Monde notes these performances remain to be confirmed by the regularly updated rankings of independent third parties.
Release plan and the road ahead
The company promised detailed results on October 27, when Mistral Large 4 is due to be released as open source. Until then, security testing will be completed, the start-up said, at a time when the sector has been rattled by incidents linked to AI models. Before the open-source release, the model can be accessed through an interface that Mistral AI says it is monitoring to prevent misuse. It will eventually become the default model for the company's assistant, Vibe, though users will still be able to choose Chinese models such as GLM, which continue to outperform in some areas.
The announcement answers suspicions that Mistral was abandoning the expensive frontier-model race after its summer 2026 investments to build its own data centres in Europe. CEO Arthur Mensch told Le Monde in September that pushing model capabilities remains the company's number one pillar, and the start-up says it is optimistic about improving performance in Mistral Large 5 and 6 over the coming year on the same technical architecture.
Lample said Large 4 was trained on 4,000 state-of-the-art chips — "two to three times fewer" than what Chinese start-ups can access, and far below the hundreds of thousands used by American leaders. Mistral aims to reach data-centre capacity of 1 gigawatt by 2030, shared between training its own models and running AI for clients, and is counting on chips coming online over the next six months following its two most recent fundraising rounds, including a €3 billion raise in early September.
Sources:
- Le Monde: Mistral AI unveils new AI model aimed at 'narrowing the gap' with top Chinese competitors
