Mistral Large 4: a 1-Trillion-Parameter Open-Weight AI Model

October 7, 2026 · Cortex Solutions

Mistral AI previewed Large 4, a roughly 1-trillion-parameter open-weight AI model, with weights due October 27. Key facts and what it means for businesses.

French company Mistral AI released a preview of Mistral Large 4 on October 6, 2026. The model has roughly one trillion parameters, handles both text and images, and comes with a promise of publicly downloadable weights, meaning companies will be able to run it on their own servers.

According to The Next Web, Mistral calls Large 4 the most powerful open-weight AI system outside China, a company claim that has not been independently verified. Internally the model goes by the nickname "Le Chonk".

Key specifications

• Size: about 1 trillion parameters in total, with around 49 billion used for any given task, which makes it more efficient to run than a dense model of the same size.
• Training: trained from scratch over about two months on roughly 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centres, using about 10 megawatts.
• Languages: trained on more than 160 languages, including all official EU languages.
• Modalities: reads text and images.
• Reinforcement learning: still ongoing at launch, and Mistral says gains have not yet started to slow.

Availability and timeline

The model has been available through Mistral's API since October 6, and the weights are scheduled for public release on October 27, 2026. In between comes a three-week real-world testing period in which developers, cybersecurity firms and government agencies get a less restricted version with additional cyber-defence features.

Customers can run the model on their own servers or use the API in a region of their choice, including an EU region covered by European data protection law. Detailed licence terms and API pricing had not been announced at the time of the preview, both of which matter for any company planning commercial use.

Early results, handle with care

Mistral published preliminary results that it says may change before the weights are released. On the DeepSWE agentic coding benchmark, Large 4 scored 62%, against 61% for GLM-5.3 and 57% for DeepSeek-V4-Pro. On the Finch/FinWorkBench finance benchmark it scored 67%. On the Harvey Legal Agent Benchmark, scores were low across all models, with Large 4 at 15%. These are self-reported numbers; independent evaluation will be the real test.

Business context

Large 4 is the first model from the €3 billion funding round Mistral closed in September 2026. The company works with more than 125 large businesses, including Airbus, ASML and HSBC, some of which helped train the model through Mistral's Forge platform. Co-founder and chief scientist Guillaume Lample said that "the cyber defense capabilities will enable enterprises and governments to defend themselves." CEO Arthur Mensch, speaking at a conference in Abu Dhabi, said: "So the narrative that Europe cannot compete is something that is not true."

Mistral plans to use Large 4 as a base for specialised models, and before the weights release it intends to publish more on training, post-training refinement, benchmarks and safety testing. Additional compute is due to come online through the first half of 2027.

What this means for Saudi businesses

Open-weight models give companies an important option: running AI inside their own infrastructure or with a local cloud provider instead of sending data to an external service. That matters for sectors handling sensitive data, such as finance, healthcare and government, and fits well with data protection and data residency expectations in the Kingdom.

It is worth staying realistic, though:

• A trillion-parameter model needs substantial multi-server compute, even if only a fraction is active per task, because all parameters must sit in memory.
• Mistral has not announced specific Arabic support, so quality on Arabic content should be tested before relying on it.
• Licence terms are not yet published and must be reviewed before commercial use.
• In many cases a smaller model tuned for a specific task is a better and cheaper fit than a giant one.

At Cortex Solutions we help companies choose the model that matches their actual need, compare open and commercial options, and build AI solutions that run in the cloud or on-premises with data security and cost in mind.

Sources

• The Next Web — https://thenextweb.com/news/mistral-releases-large-4-a-1-trillion-parameter-open-weight-ai-model