Apple HubAIPhones & PlansLaptops & PCsAudioCameras & Creator GearSmart HomeTVs & StreamingPrinters & Office
TRENDINGAI Models

Mistral launches Large 4, a 1-trillion-parameter preview model, with weights promised by end of October

Mistral released a public preview of Mistral Large 4 on October 6: a 1 trillion parameter multimodal model with 49 billion active parameters. Weights are not out yet; Mistral says they follow by the end of the month.

Key points

  • Mistral released a public preview of Mistral Large 4 on October 6: a 1 trillion parameter multimodal model with 49 billion active parameters. Weights are not out yet; Mistral says they follow by the end of the month.

Mistral AI released a public preview of Mistral Large 4 (ML4, nicknamed "le Chonk") on October 6, 2026. It is not an open-weight release yet.

What Mistral's announcement says:

- ML4 has 1 trillion parameters, 49 billion of them active, and is natively multimodal.

- The preview API is available now on Mistral Studio. Mistral says the weights will be released "by the end of the month."

- Until then, Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber capabilities.

- It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe. The preview is served from the same infrastructure, and Mistral says it will also run in other regions worldwide, including a European deployment it operates end to end.

- Mistral says a large share of the training data was multilingual, covering more than 160 languages.

Mistral's performance claims (company figures, not independently verified by us):

- Mistral calls ML4 state of the art among open models on cybersecurity, finance and law workloads, and says it beats open-weight models from the US and Europe.

- On the Artificial Analysis Cyber Index, Mistral says ML4 ranks in the top five models overall. It reports 82% on one test that asks a model to reproduce and then patch a real vulnerability, and 93% on Cybench.

- Mistral says some closed models score near zero on that test because they refuse the task. We did not check that claim.

- Mistral's page says benchmark charts for coding, agentic and finance tasks are published, and that further architecture details and benchmarks will come before the weights.

What TechCrunch adds: it reports the model is reachable for now only through a public guardrail endpoint, and quotes Mistral VP Science Pierre Stock on the plan to open the weights after safety testing. TechCrunch gives the GPU count as about 4,000; we use Mistral's own figure of 3,800. It also notes benchmark results were still pending at publication.

What to watch: whether the weights actually ship by the end of October, and whether independent evaluations match Mistral's numbers.

Sources: Mistral, "Introducing Mistral Large 4" (October 6, 2026); TechCrunch, October 6, 2026.

Primary sources