Mistral opens a preview of Large 4, a 1 trillion parameter open-weight model

Mistral on 6 October 2026 opened a public preview of Mistral Large 4, nicknamed "le Chonk," in a company post introducing the model. Developers can call it today through a preview API on Mistral Studio. The open weights are not out yet. Mistral says they will ship by the end of October, and The Next Web reports a date of 27 October.

The company describes a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 49 billion active per token. Its model documentation lists slightly different figures: 1.05 trillion total, 52 billion active, a 1.6 billion parameter vision encoder and a 1 million token context window. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the training data spans more than 160 languages.

Until the weights land, Mistral is red-teaming the model with cybersecurity firms, vetted partners and state authorities. Those groups get a version with reduced moderation and expanded cyber capabilities. Mistral pitches security work as the headline use: it reports a top five rank on the Artificial Analysis Cyber Index and 82% on a test that asks a model to reproduce and then patch a real software flaw. SiliconANGLE reports the same cyber results and notes Mistral kept training after this checkpoint, so larger versions may follow.

Key facts

• Primary: Mistral announcement and Large 4 docs, 6 October 2026
• Independent: The Next Web and SiliconANGLE, 6 October 2026
• Access: public preview API now; open weights promised by end of October
• Size: about 1 trillion total parameters; 49B or 52B active depending on the source
• Preview pricing in the docs: $0.68 per million input tokens and $2.09 per million output tokens
• Coding claims: 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, per Mistral

How to read it

Every benchmark number here comes from Mistral, and the company says the results are preliminary and may change before the weights ship. The active parameter count already differs between the blog post and the docs. The real test is the weight release: once outside teams can download and run the model, its claims on cyber and coding can be checked. Until then this is a hosted preview, not an open model in hand.

Independence

Frontier Flux is not affiliated with Mistral, NVIDIA, The Next Web or SiliconANGLE. Details come from Mistral's post and docs plus the two outlet reports, not from our own testing.