Mistral Large 4: a 1 trillion parameter model with open weights still three weeks away

What is Mistral Large 4? It is a mixture-of-experts model with about 1 trillion total parameters and 49 billion active on any one token, announced on October 6, 2026. It is natively multimodal, handles more than 160 languages including every official EU language, and is open to developers today only as a public preview through Mistral Studio. The part that makes it a race story is the promise attached to it: Mistral says the weights will follow at the end of October, and that the model will then be the strongest open-weight system outside China.

That promise is worth separating from the release. A model you can call through an API is a product. A model whose weights you can download, inspect and fine-tune is a different kind of asset, and nobody outside Mistral can check the second claim yet.

What Mistral actually published

The announcement lists the following, all self-reported:

• Size: about 1 trillion total parameters, 49 billion active, a hybrid instruct and reasoning design.
• Compute: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. Euronews, covering the launch the same day, gives the figure as 4,000 GPUs and says training took two months from scratch. The two numbers differ, and we use Mistral's.
• Price: $1.36 per million input tokens and $4.18 per million output tokens during the preview.
• Cyber: a ranking among the top five models globally on the Artificial Analysis Cyber Index, 82% on a vulnerability reproduction test and 93% on Cybench.
• Coding and agents: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, a Coding Agent Index score of 49.8%, and 59.9% on AutomationBench.
• Vision: 42% on Dense 200, against 41% for OpenAI's GPT-6 Astra.

Two things stand out. First, the one-point gap over GPT-6 Astra on Dense 200 is inside the noise a single benchmark can produce, so it tells you almost nothing. Second, the cyber results are the headline, and cyber is exactly the area where labs have started gating access to their strongest models. Mistral is choosing to ship strong cyber capability into an open-weight release.

Why the open-weight part matters for the race

Open weights change who can compete. Our recent coverage of Reflection AI's Beam made the same point from the American side: an announced model with weights weeks away is a press cycle, not a release. Mistral is in the same position today, one day later.

There is a business logic too. Epoch AI's October 6 report on how Chinese AI companies make money found that releasing weights erodes pricing power: when Z.ai released the weights of GLM 5.3 Flash on OpenRouter, its share of that model's traffic fell from 88% to 22% within 20 days as cheaper hosts appeared. Mistral raised €3 billion at a €21 billion valuation, according to Euronews, and sells on sovereignty, not on being the cheapest host. Our earlier piece on GLM 5.3's cyber capabilities covered the Chinese side of the same open-weight race.

The sovereignty pitch is simple. A European government or bank that wants frontier-class cyber and coding capability, hosted in Europe, under European law, has few choices today. Mistral wants to be the default.

What nobody has verified

Several claims are open, and an honest reading keeps them open.

The ranking. "Strongest open-weight model outside China" is a comparison against models Mistral chose. It will only mean something when independent evaluators run the same tasks on the released weights.

The license. The announcement text we opened did not state one. Open weights under a restrictive license, such as a revenue-sharing clause of the kind Moonshot attached to Kimi K3 according to Epoch, are a different thing from open weights under Apache 2.0.

The safety posture. A model that scores in the top five on a cyber index, released with downloadable weights, cannot be recalled. Anthropic now releases its strongest cyber capability only to vetted users, as we report in our piece on its Cyber Verification Program. Mistral has said that cyber defense is the use case it wants. It has not, in the material we opened, said what stops a downloaded copy being used for attack. We did not interview anyone for this piece.

The compute count. The gap between 3,800 and 4,000 GPUs is small, but it is the kind of discrepancy that shows how thin independent verification is in the first 48 hours after a launch.

How to read the benchmark list

The scores deserve a careful reading. Cybench and the Artificial Analysis Cyber Index are third-party instruments, but the numbers in the announcement are Mistral's own reports of how its model did, and the preview endpoint is the only way to reproduce them today. The coding scores sit on newer benchmarks, DeepSWE v1.1 and SWE-Atlas-QnA, that most readers will not have seen before, so there is little history to say what 61.7% means against last quarter's leaders. Our explainer on what AI coding benchmarks actually measure is a good primer on why that matters. The safest reading is that Mistral is in the conversation on cyber and coding, which is a real claim for a European lab, and that the ranking waits for outside testing.

What to do with this

If you build on European infrastructure, test the preview on your own tasks now. The price is low enough that a week of testing costs little, and your own tasks matter more than any index.

If you track the race, hold two dates. The first is the end of October, when the weights are meant to appear. The second is whatever day Artificial Analysis or Epoch publish independent scores on them. Until both have passed, Large 4 is a credible announcement from a well-funded lab, not yet a verified leader.

Our position: the size, price and compute figures are probably accurate because they are cheap to state and easy to catch if wrong. The ranking is a forecast. Do not repeat it as a fact.