Mistral releases Large 4 and opens API for preview; weights planned for October 27 public release
BIBIBI
AT A GLANCE
Mistral releases Large 4 and opens API for preview, but the claims of being “the strongest in Europe and the United States” and the benchmark advantages come mainly from self-testing.
Article
Mistral Large 4 Release: 1050000000000 parameters, claims to have built the strongest open-source model in Europe and the United States
Dongcha Beating AI breaking news: French AI company Mistral has released its new flagship model, Mistral Large 4, internally nicknamed “Le Chonk.” It uses the MoE architecture, with total parameters reaching 1050000000000, but activates only 49000000000 parameters for each inference. The model natively supports image-and-text input, with a maximum context length of 1000000 token。API is now available for preview, and the model weights are planned for public release in October 27.
Mistral says that Large 4 is currently the strongest-performing open-weight model developed in the United States and Europe overall. Its reported DeepSWE v1.1 software engineering benchmark score is 62%, while its comparison table shows GLM-5.3 at 61%、DeepSeek V4 Pro at 57%、Qwen 3.8 Max at 51%. On the Finch financial benchmark, it scored 67%, tying DeepSeek V4 Pro; on the satellite-image target localization benchmark DIOR-RSVG it scored 73%, higher than GPT-6 Astra’s 68%.
However, these results currently come mainly from Mistral’s own testing and are not based on uniform criteria. When Zhipu released GLM-5.3, it reported a score of v1.1 on the same DeepSWE 66.9%，Kimi K3 reached 67.5%, both higher than Large 4’s 62%. The testing configurations and execution methods used by different teams may vary.
Large 4 was trained from scratch. Mistral used approximately 4000 Nvidia Grace Blackwell GPU and trained it for approximately 2 months at its own data center in Europe. Mistral also promotes “European sovereignty” as a selling point: from training and API services to future self-hosted deployment, everything can remain on European infrastructure.
French AI company Mistral releases its flagship model, Mistral Large 4, internally nicknamed “Le Chonk.”
02
Large 4 uses MoE architecture, with total parameters of 1050000000000 parameters activated for each inference, 49000000000.
03
The model natively supports image-and-text input, with a maximum context length of 1000000 token。
04
API is now available for preview, and the model weights are planned for public release in October 27.
05
Mistral says that Large 4 is currently the strongest-performing open-weight model developed in the United States and Europe overall.
06
Mistral reports that Large 4 scored v1.1 on DeepSWE DIOR-RSVG, Finch, and 62%、67%respectively, and 73%。
07
Reports indicate that these results mainly come from Mistral’s own testing and are not based on uniform criteria.
08
Zhipu reports that GLM-5.3 scored v1.1 on DeepSWE 66.9%，Kimi K3 at 67.5%。
09
Mistral used approximately 4000 Nvidia Grace Blackwell GPU and trained it from scratch for approximately 2 months at its own data center in Europe.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is an image-and-text model with a very large parameter scale.MoE This means that only part of the model’s parameters are used during inference, so its total parameter count is 1050000000000, while each inference activates only 49000000000. It can now be tried through the API preview, while the model weights remain at the planned-release stage.
Why it matters to readers
Large 4 It also emphasizes its very large parameter scale, image-and-text input, 1000000 token context, plans for open weights, and European infrastructure. If the weights are released as planned, users will have another option for trying the model and deploying it independently; however, the existing performance rankings cannot be directly treated as conclusions from uniform testing.
Beginners should pay attention to API how easy it is to try, whether the model weights are released as planned, and its actual performance under uniform testing conditions; they should not judge the model solely by claims that it is “the strongest” or by a single set of scores.
Risks and unknowns
“The strongest-performing open-weight model in the United States and Europe overall” is Mistral’s own claim and still lacks validation under uniform criteria.
Different teams may use different testing configurations and execution methods, so existing DeepSWE v1.1 scores cannot be directly compared across models.
The model weights are only planned for public release in October 27; there is no confirmation that the planned release will be completed.
Related Developments
Loading event timeline…
The evidence does not provide API the preview’s actual usability, cost, or independently tested results.
Related concepts
MoE
This term is not in the glossary yet. Browse related concepts in the glossary.
