Mistral Large 4: 1T-Parameter Open-Weight Preview, Pricing and Specs

What Mistral Large 4 is

Mistral Large 4 is Mistral's largest model, launched as a public preview on October 6, 2026. Mistral calls it ML4 for short and, in its own words, "very officially: le Chonk". It is a natively multimodal model with 1 trillion total parameters and 52 billion active. The preview API is available on Mistral Studio, and Mistral states it will release the weights by the end of October 2026; until then it is red-teaming the model with cybersecurity partners, vetted partners and state authorities. Mistral says it will publish further details on the architecture and post-training as it works toward the weight release.

Pricing and specifications

Mistral lists the preview API at $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. When checked on 12 October 2026, Mistral's model documentation also showed a sale price of $0.68 input, $0.07 cached input and $2.09 output per million tokens. The documentation lists a 1M token context window and the model id mistral-large-4-0, and describes a 1.6 billion parameter vision encoder.

Training

Mistral states ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe, and that the public preview is served on the same infrastructure. Its training data spanned more than 160 languages, including every official language of the European Union. Mistral says the model uses the same training, customisation and reinforcement-learning environment it offers customers through Mistral Forge.

Benchmarks, and who ran them

These are Mistral's reported results. On coding it reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, combining to a Coding Agent Index of 49.8%. In cybersecurity it reports 93% of Cybench's 40 challenges solved and 82% on a test that asks a model to reproduce and then patch a real vulnerability. On AutomationBench, 657 business workflows, it reports 59.9%. In a blind human evaluation of coding quality run with Surge AI on a 1 to 5 scale, Mistral reports ML4 Preview at 3.74, behind Claude Opus 5 at 4.22 and ahead of GLM-5.3 at 3.60, Kimi K3 at 3.59 and GLM-5.2 at 3.40. On visual grounding it reports 42% on Dense 200 against 41% for GPT-6 Astra.

Other results Mistral reports

Mistral states its Coding Agent Index score of 49.8% places ML4 ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, and that on AutomationBench it scores ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro. It reports that third-party evaluator vals.ai found the model ahead of GPT-6 Astra on representative legal and financial tasks, that it outperforms all open-source models on Harvey's Legal Agent benchmark, and that it is state of the art among open-weight models on SciCode-Verified. All of these are Mistral's statements about its own preview.

Deployment

Mistral states the model will be available across multiple regions, including a European deployment it operates end to end under European law, and that once the weights are released it can run on private cloud or on-premise. Mistral describes cybersecurity as a focus, noting that several closed models score near zero on its vulnerability-reproduction test because they refuse the task; that is Mistral's account of its own test.

Official Resources