Mistral releases 1 trillion-parameter Large 4 model under a restricted open-weight license

• Mistral Large 4 is available in FP8 and NVFP4 versions, with roughly 1 trillion total parameters, about 49 billion active parameters per token and a 1 million-token context window. ^[1]
• The license permits commercial use, modification and fine-tuning, but companies offering the model as a service need a separate commercial license from Mistral. ^[2]
• Artificial Analysis currently lists the released model at 26 on its Intelligence Index, with API pricing of $0.70 per million input tokens and $2.10 per million output tokens. ^[3]
• Mistral reports a 51% overall score on the Artificial Analysis Cyber Index. The 82% figure often cited for vulnerability reproduction and patching comes from the earlier preview evaluation and should not be presented as the released model’s overall cyber score. ^[4]

Mistral AI has released the weights for Mistral Large 4, making its largest model available for download and self-hosting after a public preview earlier in October. The French company is positioning the model for coding agents, enterprise workloads and defensive cybersecurity. [5]

Large 4 is a multimodal mixture-of-experts model with roughly 1 trillion total parameters. Mistral’s model card describes about 49 billion active parameters per token, or 52 billion when embeddings and output layers are included. It supports text and image input, a 1 million-token context window, function calling and three reasoning settings. Mistral says it was trained on 3,800 NVIDIA Grace Blackwell GPUs in European data centers. [1]

Downloadable weights, with a service restriction #

The release includes official FP8 and NVFP4 checkpoints, while community quantizations are available for llama.cpp and related tools. Hugging Face lists support for Transformers, vLLM, SGLang and llama.cpp. Mistral also makes the model available through its API, Microsoft Foundry and NVIDIA-hosted infrastructure. [1][2]

The model is distributed under Mistral’s Ecosystem Open-Weight License, which the company describes as a modified MIT license. It permits commercial and noncommercial use, modification and fine-tuning, but requires a separate commercial license for a business that operates a Model-as-a-Service offering. Artificial Analysis correspondingly labels the model’s open-weight status as “Commercial Use Restricted.” [2][3]

Mistral’s API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Artificial Analysis lists the same standard input and output prices and estimates an average cost of $2.06 for one task in its Intelligence Index evaluation. Those figures describe hosted inference; they do not establish the cost of running the downloaded model on a company’s own hardware. [2][3]

Coding performance improved after the preview #

Mistral says Large 4 scores 48% on Artificial Analysis’ Coding Agent Index, effectively tying GLM-5.3 while using 47% fewer tokens per task. The company also reports 19.7% on Cognition’s private FrontierCode 2.0 benchmark and 41% on Terminal-Bench 4.0, which it says was a 13-point improvement over the public preview. [2]

Artificial Analysis currently gives the released model a score of 26 on its Intelligence Index, ranking it ninth among 21 models on the page. The index combines nine evaluations, including FrontierCode 2.0, Terminal-Bench 4.0, AutomationBench and GDP.pdf. Artificial Analysis describes the model as above the comparison median, but not a leader on the aggregate index. [3]

The distinction between checkpoints matters. Hugging Face’s launch article lists 28.3% on Terminal-Bench 4 and a 49.8% combined Coding Agent Index score, while Mistral’s release post reports the newer 41% Terminal-Bench result and 48% Coding Agent Index score. Mistral has not yet published the full technical report promised for the release, so the precise evaluation settings and checkpoint differences are not fully documented. [1][2]

Cybersecurity claims need narrower wording #

Mistral reports a 51% score on the Artificial Analysis Cyber Index, which measures vulnerability discovery, reproduction and patching. It also says the model refused 96.6% of malicious cyber requests in its safety evaluation. These are company-reported results cited from an independent evaluation framework, not an independent audit of Mistral’s full release announcement. [2]

The 82% figure repeated in the Hugging Face article refers to a specific CyberGym reproduction-and-patching test from the public-preview evaluation. Artificial Analysis’ October 6 preview article reported that score alongside a 50% overall Cyber Index result. The released model’s current Mistral announcement reports 51% on the overall index but does not present 82% as the overall score. The figures should therefore not be collapsed into a single claim that Large 4 scored 82% on cybersecurity. [1][2][4]

Mistral is offering a specialized cybersecurity variant to selected design partners, including Orange Cyberdefense, for penetration testing, vulnerability assessment and red-teaming. The company says access is restricted to authorized systems and subject to contractual safeguards. For the downloadable general model, self-hosting gives organizations greater control over data and deployment, but it does not eliminate the need for isolation, permissions and human review when the model can generate exploit code or modify production systems. [2][6]

Companies mentioned #

Further sources #

The stories that matter, in one email. Free — unsubscribe anytime.