Reboot Germany

Mistral brings Europe closer to the AI leaders

Mistral brings Europe closer to the AI leaders. Germany has a smaller model and a planned merger with Cohere in Canada.

By Alec Chizhik October 6, 2026 4 min read

Mistral Large 4 has been in public preview since 6 October, after training in Mistral’s own European data centres. Its 38 points in the Intelligence Index make it the highest-ranked model outside the US and China. Open weights are due at the end of October. Aleph Alpha’s German counterpart is smaller.

Key takeaways

Mistral Large 4 is the strongest model outside the US and China. The preview scores 38 points in the Intelligence Index; its predecessors scored 14 and 9 points.
It remains more than 19 points behind the leader. Claude Opus 5.5 leads with 58 points. Among open-weight models, seven Chinese models would rank ahead of Mistral.
Open weights remain a promise until they are released. Until the end of October, the model runs only through the API and is classified as proprietary.

A jump in the index

Since 6 October, Mistral Large 4 has been available in public preview through the API in Mistral Studio. In the Intelligence Index from benchmark provider Artificial Analysis, a score based on numerous test tasks, Mistral jumps from 9 to 38 points. This score makes Mistral Large 4 the highest-ranked model outside the US and China, ahead of countries such as South Korea and the United Arab Emirates.

What is an open-weight model? Weights are the parameters that an AI model learns during training and that determine its behaviour. With an open-weight model, the provider makes these parameters available to download, allowing companies to run the model on their own hardware or in their own cloud. Mistral Large 4 is classified as proprietary until the announced release. Aleph Alpha’s Kolibri is already available under Apache 2.0.

The gap to the leaders

The jump significantly narrows Mistral’s gap to the leaders, but does not close it. The table shows where the model stands in the index.

Model	Provider	Origin	Points
Claude Opus 5.5	Anthropic	USA	58
Claude Sonnet 5.5	Anthropic	USA	56
Claude Fable 5.1	Anthropic	USA	53
GPT-6 Astra	OpenAI	USA	53
Gemini 4 Argon	Google	USA	53
MiMo-V2.6-Pro	Xiaomi	China	46
GLM-5.3	Z.ai	China	45
Kimi K3	Moonshot AI	China	44
DeepSeek V4.1 Flash	DeepSeek	China	39
Mistral Large 4 (preview)	Mistral AI	France	38
GPT-6 Luna	OpenAI	USA	38
Nemotron 3 Ultra	Nvidia	USA	23
Mistral Medium 3.5	Mistral AI	France	14
Mistral Large 3	Mistral AI	France	9

Source: Artificial Analysis Intelligence Index v4.3.2, as of 6 October 2026; rounded scores. Assessment based on Artificial Analysis, Trending Topics and OfficeChai

The gap to the leader is around 20 points. Among open-weight models, Mistral Large 4 would rank eighth, with all seven models ahead of it coming from China.

Pricing and operation in Europe

Billing is based on tokens, the text units into which a model breaks down inputs and outputs. Mistral Large 4 is considered highly verbose. At list price, an index task costs around four times as much as it does with cheaper alternatives.

Metric	Mistral Large 4
Parameters	around 1 trillion, of which 49 billion are active
Training	3,800 Nvidia Grace Blackwell GPUs in its own data centres in Europe
Languages	more than 160
Input list price	around 1.21 euros per million tokens
Output list price	around 3.71 euros per million tokens
Launch discount	50 per cent
Cost per index task	around 1 euro (GLM-5.3-Flash: around 0.22 euros)
Speed	116 tokens per second
Open weights	announced for the end of October

Source: Mistral AI, Artificial Analysis, Huggingnews; euro amounts converted at the ECB exchange rate of 6 October 2026

According to the manufacturer, it should become possible to run the model in a private cloud or on premises, meaning a cloud environment reserved for one company or the company’s own data centre. Operation on European infrastructure is intended to be possible independently of other digital service providers and under European law.

Mistral says the weights are due at the end of October. Until then, the model runs only through the API and is classified as proprietary in the Intelligence Index. Open weights remain a promise until they are released.

Funding and investors

Training was financed by a 3-billion-euro Series D, which the company describes as the largest equity funding round ever raised by a European technology company. Following the Series D, the three-year-old company is valued at more than 21 billion euros. It intends to work closely with investors such as ASML and Samsung.

The German counterpart

On 3 October, Aleph Alpha released Kolibri as an open model under the Apache 2.0 licence. Kolibri also uses a mixture-of-experts architecture, has around 78 billion parameters and was trained on infrastructure in Germany and Finland.

France therefore leads the country comparison for the strongest model outside the US and China. The German counterpart is a smaller open model from a lab that is set to merge with Cohere from Canada. On 16 September, Cohere from Canada and Aleph Alpha signed the merger agreement . The combined company is expected to operate under the Cohere brand. Regulatory approval is still pending.

Frequently Asked Questions

How can Mistral Large 4 be used today?

The preview runs through the API in Mistral Studio. It is also accessible through the OpenRouter platform.

Which languages and inputs does the model support?

Mistral Large 4 covers more than 160 languages and was built to be natively multimodal. It processes up to 100 images per request.

How does Mistral Large 4 perform on cyber tasks?

The model scores 50 points in Artificial Analysis’s Cyber Index.