Mistral AI Unveils Mistral Large 4, a 1.05-Trillion Parameter Multimodal Model

A Giant Among Open Models

Mistral AI has announced the release of Mistral Large 4, internally dubbed "Le Chonk," as a public preview on October 6, 2026 ¹. The model is a granular Mixture of Experts (MoE) architecture with a staggering 1.05 trillion total parameters, though only 49 billion are activated per token . It also features a 1.6-billion-parameter vision encoder and supports a 1 million token context window, making it one of the most capable multimodal models available today .

What makes this architecture efficient is that only about 4.7% of the total weights activate for each token processed. This means the model can serve at mid-tier pricing while carrying the knowledge capacity of a trillion-parameter class system ¹. The full 1.05 trillion parameters still need to reside in memory, but the lower active count controls compute costs rather than hardware bills .

Where It Trains and What It Knows

Mistral Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral's own European datacenters ¹. The training data covered more than 160 languages, including every official language of the European Union . The model natively accepts image input alongside text, marking it as a true multimodal system rather than a text model with an add-on vision component .

The API is now live at a preview price of $1.36 per million input tokens and $4.18 per million output tokens ¹. However, the full model weights will not be available until the end of October 2026, meaning self-hosting is not yet possible . The license has also not been announced .

Strong Performance in Cybersecurity

One of the model's standout areas is cybersecurity. Mistral AI reports that Mistral Large 4 scored 93% on Cybench and 82% on CyberGym-E2E, placing it in the global top 5 on the Artificial Analysis Cyber Index ¹. More notably, Mistral claims several closed frontier models score near zero on CyberGym-E2E because they refuse to attempt the tasks at all . Reproducing vulnerabilities is standard defensive cybersecurity work, and provider-level refusals can block that process .

In agentic coding benchmarks, the model achieved 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.0, earning a combined Coding Agent Index score of 49.8 ¹. Mistral notes these evaluations were run privately before the benchmarks went public, so they are not yet independently reproducible . In blind human evaluations by Surge AI, annotators rated Mistral Large 4 Preview at 3.74 out of 5, placing it second among five models tested, behind Claude Opus 5 which scored 4.22 .

On the safety front, the model resisted 93.3% of attacks on Lakera's B3 benchmark and scored 1.691 out of 2.0 on KORABench ¹.

Rivals in the Arena

When compared to other open-weight models like DeepSeek V4 Pro, Kimi K3, and GLM-5.3, Mistral Large 4 holds its own with the largest active parameter count among its peers and native image understanding ¹. The model supports function calling, structured outputs, document Q&A, batching, and Agents and Conversations endpoints . Cached input is priced at just $0.14 per million tokens, which significantly improves the economics of running long-context agent workflows .

Full details and documentation are available through the Mistral Large 4 announcement and model docs ¹.

Key facts

• Mistral Large 4 has 1.05 trillion total parameters with 49 billion active per token
• It features a 1 million token context window and native image input
• Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters
• Scores 93% on Cybench and 82% on CyberGym-E2E in cybersecurity benchmarks
• API is live now; full weights expected by end of October 2026

Sources

1. Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model — marktechpost.com