Mistral Large 4: Europe's 1-Trillion-Parameter Open-Weight AI Model, Explained

Europe's biggest AI company has made its boldest move yet. On October 6, Paris-based Mistral AI opened a public preview of Mistral Large 4 — a model with around a trillion parameters that the company promises to release as open weights, meaning anyone will be able to download and run it.

What is Mistral Large 4?

Mistral Large 4 is a general-purpose model that understands both text and images. It uses a Mixture-of-Experts design: instead of using the whole model for every request, it activates only the "experts" it needs, which keeps it faster and cheaper to run than its size suggests.

Mistral's own pages give slightly different numbers for the parameter counts, and at least one independent tracker lists a smaller context window, so the details may still change.

How good is it?

Mistral says Large 4 is competitive with the strongest open models and leads open models on enterprise work like cybersecurity, finance and law. The company reports scores including 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA for coding, and says the model solves 93% of the challenges in the Cybench security test. These are Mistral's own results — independent testing will tell us more.

What independent tests say

The only third-party check we found is Vals AI. As of October 14 it lists Large 4 at 48.1% on the Vals Index, 35th of 48 models. Its best result is 9th of 79 on Harvey's Legal Agent Benchmark (15.8%), and it ranks 24th of 78 on the Finance Agent v2 test. That fits Mistral's focus on enterprise work, but it also shows that the model is not at the top of the overall rankings. For comparison, the leader on the same index, Google's Gemini 4 Argon, scores 68.9%. Large 4 is a leading open model, not the strongest model overall.

Price

For its first two weeks, Mistral is offering the preview at 50% off — about $0.68 per million input tokens and $2.09 per million output tokens, compared with the regular $1.36 and $4.18.

Why aren't the weights out yet?

Because of the model's strong cybersecurity abilities, Mistral is first red-teaming it with cybersecurity experts, vetted partners and government authorities before releasing the weights. Companies that need to run it on their own servers now can do so under a separate commercial licence.

Until the weights are actually published, "open-weight" is a promise rather than a fact — and one independent benchmark site currently marks the preview as not open-weight for that reason.

Who should care

• Companies with privacy or sovereignty rules that want to run a strong model on their own servers or private cloud.
• Developers who want to avoid lock-in and may fine-tune the model once the weights are public.
• Everyone else can wait: a trillion-parameter model needs serious hardware, and Mistral has not said what is needed to self-host it.

Why it matters

Mistral markets itself as the leader in "sovereign AI" — AI that governments and companies, especially in Europe, can control and run themselves rather than relying on U.S. or Chinese providers. A trillion-parameter open model would be a major step toward that goal and a direct challenge to the top open models from China.

The reality check

• Benchmark numbers are Mistral's. Most published scores come from the company, and independent rankings so far place the model mid-table.
• The weights aren't out. Mistral says they will arrive by the end of October, and the licence has not been named.
• Cyber capability cuts both ways. Mistral is red-teaming the model with security experts and state authorities first, and its cyber-focused deployment will have reduced moderation.

Do you think open-weight models can keep up with closed ones like GPT and Claude?

Sources: Mistral AI, Mistral docs, Let's Data Science, VKTR, Vals AI. Image: Mistral AI.