Mistral released the weights for ML4, its trillion-parameter multimodal model, on October 29, 2026, after a public preview that began October 62. The open-weight drop gives businesses and governments a system they can download, customize and host without relying on Mistral's API.
Mistral says it trained ML4 from scratch on European infrastructure, with traceable origins from development through release. The company positions the model at the center of its sovereign-AI offering, letting organizations inspect, modify and run the system themselves.
Benchmark claims and coding performance
Mistral reports that ML4 leads open-weight models on FrontierCode 2.0, where it scored 19.7%, and on Terminal-Bench 4.0, where it reached 41%. Additional scores Mistral cites include 84.5% on BrowseComp, 73.3% on Finch's finance workflows, 51% on the Artificial Analysis Cyber Index and 48% on the Artificial Analysis Coding Agent Index. Attack resistance on the B3 Agent Security Benchmark rose to 97.1% from 93.3% during preview.
ANALYSIS The FrontierCode and Terminal-Bench leads are self-reported and scoped to open-weight models; how ML4 compares to closed frontier systems on those same benchmarks is not addressed in the available claims.
Cybersecurity and access controls
Mistral reports a 96.6% average refusal rate for malicious cybersecurity requests. The company is giving enterprise design partners, including Orange Cyberdefense, preview access to a specialized ML4 variant for penetration testing, vulnerability assessment and red-teaming. Access to that variant requires vetting and contractual safeguards, and tests are limited to owned or explicitly authorized systems in controlled environments. Mistral Vice President of Science Pierre Stock said open weights give organizations access to defensive AI that cannot be cut off by an outside provider.
Mistral is updating its cybersecurity guardrails and working with customers to customize models around proprietary security data, detection logic and threat environments. Defined security tasks can be optimized into smaller, specialized variants.
Pricing, licensing and deployment
API pricing for ML4 is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Vector Wire's prior coverage noted those rates represent roughly a halving from the $1.36/$4.18 input/output pricing listed on October 21[3]. The Mistral Ecosystem Open-Weight license permits free use, modification, fine-tuning and commercialization, though businesses offering Model-as-a-Service must obtain a separate commercial license.
ML4 offers three reasoning modes — low, medium and high — with high as the default. Deployment options include on-premise hosting and Microsoft Foundry on Azure, and enterprises can tailor the model using Mistral Forge and build applications through Mistral Studio. Developers can download the weights from Hugging Face and serve them using vLLM or SGLang; NVIDIA also offers hosted endpoints and an optimized NVIDIA NIM container.
La Banque Postale described the release as reinforcing its long-term partnership with Mistral.