Mistral releases Large 4 for anyone to download and run

Mistral AI released the full set of files behind Mistral Large 4, its newest and largest artificial intelligence model. The company posted them on October 29, 2026, to the Hugging Face developer platform and through its own programming interface. Read more: Mistral releases Large 4 for anyone to download and run

The release makes Large 4 what is known as an open-weight model. So the files needed to run it are free for any business or government to download. They can then operate it on their own computers, instead of reaching it through Mistral's website or paid service.

Mistral showed Large 4 on October 6, 2026, as a limited preview reached through its own system, according to Decrypt. The model holds roughly one trillion internal settings. But only about 49 billion activate for any single response, which Mistral says keeps the cost of running it lower.

Mistral charges $1.36 for every million tokens of text sent into the model. It charges $4.18 for every million tokens the model returns, according to Decrypt's review of the company's published rates.

The input price is about a third of what Anthropic charges for Claude Opus 5.5. The output price, though, is about a fifth, per Decrypt. Against OpenAI's GPT-6 Astra, the input price is roughly a seventh and the output price about a twelfth.

Related: OpenAI reportedly plans to buy Google's AI chips

Pitched as cyber defense

Mistral is marketing Large 4 directly at cybersecurity and defense buyers. Pierre Stock, the company's vice president of science, told TechCrunch that the model runs on an organization's own servers. That means access to it "can't be cut off by definition because the model is open-weighted," he said.

He added that security teams could use it to identify and respond to major incidents in real time.

Related: OpenAI epilepsy lawsuit says ChatGPT coached patient off medication before death

Chief executive Arthur Mensch framed the release in broader terms. He said artificial intelligence should "be a tool for empowerment, not dominance," according to TechCrunch. Still, the company argues that businesses and governments should not have to depend on outside AI providers. It says that is especially true for capabilities that touch their own security.

Mistral says Large 4 performs about as well as GLM-5.3, a Chinese open-weight model previously seen as among the strongest. The company also says Large 4 uses 47 percent fewer tokens per task. It also says the model ranks above every other open-weight model on the Coding Agent Index, a benchmark published by the research group Artificial Analysis.

Those figures come from Mistral's own testing. Independent researchers, however, have not yet published results checking them.

The weights for Mistral Large 4 ship under a new license. Mistral has not yet published the license in full. That marks a departure from the fully open Apache license it used for Large 3, its previous model, according to The New Stack.