Is Mistral Large 4 the Best Open Model Outside China?

On preview scores, yes. Mistral Large 4 scored 38 on the independent Artificial Analysis Intelligence Index, higher than any open model built outside China. Seven Chinese open models score higher.

Mistral launched it as a paid preview on October 6 and has promised the weights by the end of the month. It told Reuters the public date is October 27. It has not said which licence the weights will carry. Until they appear, Artificial Analysis lists the preview as proprietary.

What is Mistral Large 4?

It is Mistral’s biggest model so far. It follows the €3 billion round in September that valued the French company at around €21 billion. Training had already started by then.

Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centres. It handles more than 160 languages, including every official EU language. Its nickname is “Le Chonk”, and chief executive Arthur Mensch joked on X that it is “actually le gros chaton”, the fat kitten.

Is Mistral Large 4 Better Than DeepSeek, Kimi and GLM?

It beats DeepSeek V4 Pro, but not China’s best. On Artificial Analysis’s index, Mistral Large 4 scored 38, ahead of DeepSeek V4 Pro on 36. DeepSeek V4.1 Flash scores 39. Seven Chinese open models score higher: Xiaomi’s MiMo-V2.6-Pro (46), Z.ai’s GLM-5.3 (45), Moonshot’s Kimi K3 (44), GLM-5.3 Flash (42), Alibaba’s Qwen3.8 2.4T and Qwen3.8-Flash-Next (both 40), and DeepSeek V4.1 Flash (39). Xiaomi’s smaller MiMo-V2.6-Flash is just behind, on around 38.

Outside China, by contrast, no open model matches it. The previous leaders were South Korea’s Motif 3, at an estimated 34, and the UAE’s K2 Horizon, at 31. The best American open model, Thinking Machines’ Inkling Small, scores 26. Mistral’s own Large 3, from December 2025, scored just 9. Reflection AI’s Beam, a US rival announced a day before Mistral’s launch, has no public weights or independent score yet.

Price widens the gap. At list price, Mistral Large 4 costs about the same as GLM-5.3, which scores seven points more. MiMo-V2.6-Pro costs a fifth as much per output token. Even so, Mistral’s docs show half price during the preview.

For Europe, the release still matters. Of the labs on our European map, only Mistral and Poolside passed our test of an open model above 100 billion parameters this year. In China, thirteen labs did. “There’s an enormous desire and an enormous need for alternative technology suppliers,” Mensch told a conference in Abu Dhabi on October 6.

Where is it Strong, and Where is it Weak?

Legal work is its standout, finance is mid-table and coding is its weak spot. Vals AI, which tested it independently, ranked it as follows:

It is also wordy. Artificial Analysis counted 200 million tokens of output to run its index, against a median of 81 million. As a result, each task costs more than the price list suggests.

Mistral’s boldest claims are on cybersecurity. It reports 93% on Cybench, a set of security challenges. It also claims 82% on patching software flaws, in a test it credits to Artificial Analysis’s Cyber Index. However, Artificial Analysis had not published those results by October 6. Separately, it found that Claude Opus 5.5 and GPT-6 Astra refuse at least 98% of such tasks. Meanwhile, Pierre Stock, Mistral’s vice-president of science, told Axios: “We’re not there yet on the frontier.” Claude Opus 5.5, a closed model, scores 58 on the Intelligence Index.

Is Mistral Large 4 Open Source?

Not yet, and Mistral has not said on what terms it will be. Its docs call the model “open-weight” but name no licence, while Artificial Analysis lists it as proprietary until the weights appear.

The licence matters, because Mistral’s terms vary. Large 3 came out under Apache 2.0, which lets anyone use it commercially. By contrast, Medium 3.5’s modified MIT licence does not cover companies with over $20 million in monthly revenue. They must ask Mistral for a commercial licence or use its hosted service. VentureBeat reported that the weights are expected under a custom Mistral licence. Until the text is out, a business cannot know whether they can run the model for free.

When Can You Download Mistral Large 4, and Can You Run It?

We read Mistral’s announcement and documentation, and the Artificial Analysis and Vals AI result pages, on October 6. The scores are for the preview.

Mistral says the reinforcement-learning run behind the preview is still in flight, so the released model may score differently. Artificial Analysis published the Intelligence Index score of 38 and the Cyber Index score of 50 the same day. We will update this page when the weights and licence are published.

Author: Akos Szima

See Also:

Who Owns Mistral AI? What the Company’s Own Filings Show

European LLMs in 2026: The Complete Map

Chinese LLMs in 2026: The Complete Map

How We Checked This:

We read Mistral’s announcement and documentation, and the Artificial Analysis and Vals AI result pages, on October 6. The scores are for the preview. Mistral says the reinforcement-learning run behind the preview is “still in flight”, so the released model may score differently. We will update this page when the weights and licence are published.

Frequently Asked Questions: