Mistral Large 4 Trails China’s Open-Weight Leaders on Artificial Analysis
Trending Topics Jakob Steinschaden ● Covered by 5 sources
Mistral’s new Large 4 is strong, but not top in open models. Seven Chinese models still beat it on Artificial Analysis.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral’s new Large 4 has had its first outside test, and the result is good for Paris and awkward for the launch pitch. On Artificial Analysis, the model scores 38.4 on the Intelligence Index. That makes it the strongest open model outside China — but not the strongest open model overall.
Seven Chinese models sit ahead of it in the open-weight ranking, led by Xiaomi’s MiMo-V2.6-Pro at 46.3, Z.ai’s GLM-5.3 at 44.8 and Moonshot AI’s Kimi K3 at 43.6. Mistral had introduced Large 4 as a model that could stand next to the best open systems anywhere. The benchmark says that claim only partly holds. It does beat every open model from the US and Europe, and it clears the models Mistral highlighted at launch, including GLM-5.2 and DeepSeek V4 Pro.
The jump is still big for Mistral itself. Medium 3.5 scored 14 points on the same index, and Large 3 scored 9. But the company is not done training the preview, so this is a moving target. The weights also haven’t been released yet, which is why Artificial Analysis still treats it as a proprietary model for now. Mistral says the weights should arrive at the end of October.
Cost and behavior paint a mixed picture. Large 4 comes in at $1.13 per task on Artificial Analysis, cheaper than GLM-5.3, Kimi K3 and Qwen3.8, which are all around $2, but pricier than more efficient Chinese models such as DeepSeek V4.1 Flash and MiMo-V2.6-Pro. It also talks a lot: about 200 million output tokens across the full test, well above the 81 million median, which is why Artificial Analysis calls it “very verbose.” Still, it’s quick, with 116 tokens per second and a first response after 1.46 seconds.
What users can do with the preview is limited. Right now it’s only available through Mistral’s own API, it can’t yet be downloaded, and the public version has tighter cyber controls. The bigger unknown is the license. Mistral talks about open weights, but not the terms. And that may decide whether Large 4 is really open in the way people mean when they say it.
My take — AI-written commentary, not fact-checked reporting
Mistral has done the hard part: build a model that belongs in the conversation. The annoying part is the usual one — “open” still depends on a license fine print that everyone pretends is a footnote until a lawyer turns up. Europe keeps wanting sovereign AI, but sovereignty with a surprise restriction clause is just brochureware with better fonts.
Read more about this at: Trending Topics