The news
Mistral AI released Mistral Large 4 in preview on October 6, 2026, a model with 1 trillion parameters that the French company plans to publish as open weights on October 27, according to The Next Web. Open weights means companies can download the model and run it on their own hardware instead of only calling it through Mistral's service.
Mistral Large 4 uses a mixture-of-experts design, in which only part of the network runs for each piece of text. SiliconANGLE reported that 49 billion parameters are active per inference, which makes the model cheaper to run than its total size suggests. It supports more than 160 languages, including all official EU languages, The Next Web reported. It accepts text and images as input.
API access began on October 6 as a public preview, SiliconANGLE reported. The Next Web said developers, cybersecurity firms and government agencies will receive a version with fewer restrictions, and that regional deployment options include EU data centers. Pricing was not disclosed in either report.
Mistral trained the model on Nvidia (NVDA) Grace Blackwell hardware: about 3,800 chips according to SiliconANGLE, or roughly 4,000 GPUs over two months according to The Next Web. On benchmarks reported by SiliconANGLE, the model scored 82% on the AA Cyber Index for patching open-source projects and edged out OpenAI's GPT-6 Astra by 1% on the Dense200 visual benchmark. SiliconANGLE said it trails frontier models on coding but beats Qwen3.8 Max and DeepSeek V4 Pro.
Guillaume Lample, Mistral's co-founder and chief scientist, told The Next Web the model's cyber defense capabilities will help enterprises and governments defend against threat actors.
The numbers
- Total parameters
- 1 trillion
- Active parameters per token
- 49 billion
- Languages supported
- 160+
- Open weights release
- October 27, 2026
- AA Cyber Index score (per SiliconANGLE)
- 82%
Why CEOs should care
For CIOs and CISOs in regulated industries, an open-weight model at this scale means a frontier-class option you can host inside your own environment. That matters where data cannot leave a jurisdiction or a private network. Ask whether your team has the GPU capacity to run a 1 trillion-parameter mixture-of-experts model, or whether a hosted EU deployment from Mistral meets your residency rules.
CFOs should look at total cost, not license cost. Open weights remove per-token fees if you self-host, but you pay for hardware, power and staff. Compare that with Mistral's API price once it is published, and with closed US models, on your real workloads.
Boards and procurement teams gain leverage. A credible European open model gives buyers an alternative to depending on a few US or Chinese providers. Test Mistral Large 4 on cybersecurity, finance or legal tasks where Mistral claims strength, and treat vendor-reported benchmark scores as a starting point, not proof.
The bigger picture
Mistral is pitching Large 4 as a European answer to both US closed models and the strong open-weight models from Chinese labs. The Next Web described the launch as a challenge to China's lead in open models, and Mistral stresses that the model was built and can be deployed in Europe.
According to SiliconANGLE, Mistral did not stop the training run after producing Large 4 and expects to release larger, more capable versions in the coming months, eventually building a series of models for specific uses. The Next Web reported that more computing capacity comes online through the first half of 2027.
What’s next
The key date is October 27, when Mistral plans to publish the weights. Watch for the license terms, published API pricing and independent benchmark results before committing production workloads.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
