Mistral AI has released the weights for Mistral Large 4, a trillion-parameter multimodal model that organizations can download, customize and operate on their own infrastructure. The release expands access beyond hosted services, with Mistral positioning the system for coding, enterprise workflows and defensive cybersecurity.
The weights became available on October 29, following a public preview that began October 6. Mistral says three additional weeks of reinforcement learning improved the model’s coding performance and reduced the number of output tokens needed for some tasks. Its reported results place Large 4 among the leading open-weight coding models, though benchmark scores measure performance under particular evaluation conditions rather than guaranteeing comparable results across production workloads.
Mistral reports a 19.7% score on Cognition’s private FrontierCode benchmark, which evaluates end-to-end code quality and whether proposed repository changes are acceptable. On Terminal-Bench, the model scored 41%, improving by 13 percentage points from the preview while using 30% fewer output tokens. It also reached 48% on the Artificial Analysis Coding Agent Index, effectively matching GLM-5.3 while consuming 47% fewer tokens per task on average.
Token consumption matters for coding agents because completing a software task can involve repeated model calls, tool use and revisions. Lower usage can reduce metered inference costs, although actual expenditure also depends on pricing, task complexity and deployment configuration. Large 4 offers low, medium and high reasoning settings, allowing developers to adjust the amount of reasoning applied to a task. High is the default.
Developers can obtain the weights through Hugging Face and serve the model using vLLM or SGLang. Hosted access is available through Mistral Studio, the Mistral API and OpenRouter. NVIDIA provides hosted endpoints and an optimized NIM container, while enterprise deployment options include on-premise hosting and Microsoft Foundry on Azure.
Mistral says the model was trained from scratch on European infrastructure, with traceable origins throughout development and release. Self-hosting gives organizations greater control over where inference runs and how the model integrates with internal systems. It also leaves operators responsible for providing sufficient computing capacity, maintaining the deployment and implementing appropriate security controls.
The release uses the Mistral Ecosystem Open-Weight license, a modified MIT license that permits free use, modification, fine-tuning and commercialization. There is an important exception: businesses offering Model-as-a-Service need a separate commercial license before commercial use. Open-weight availability therefore should not be read as unrestricted permission for every business model.
Cybersecurity is another focus. Mistral reports a 51% score on the Artificial Analysis Cyber Index, which covers vulnerability discovery, reproduction, validation and patching, alongside a 96.6% average refusal rate for malicious cybersecurity requests. Those are reported evaluation results, not a guarantee that the model will reject every harmful request or perform every defensive task reliably.
Enterprise design partners, including Orange Cyberdefense, are receiving preview access to a specialized variant for penetration testing, vulnerability assessment and red-teaming. Access requires vetting and contractual safeguards, with testing restricted to owned or explicitly authorized systems in controlled environments. For the standard API, pricing is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.
Sources: qz.com, The Cryptonomist