Mistral AI Releases Mistral Large 4 Multimodal Model

The model's standout achievement lies in specialized enterprise workloads. In visual grounding tasks, ML4 reportedly surpasses several leading closed models, making it a preferred choice for industries requiring high precision. In sectors such as cybersecurity, finance, and law, the model has shown state-of-the-art accuracy in reasoning through complex, multi-step regulatory and technical queries.

To ensure safety and robustness, the model is currently undergoing intensive red-teaming in real-world settings. This process involves collaboration with cybersecurity leaders, vetted partners, and state authorities. These partners are accessing the model with reduced moderation and expanded cyber capabilities to stress-test its limits before the full weight release.

• Superior performance in coding (HumanEval) and agentic workflows
• State-of-the-art visual grounding capabilities
• Leading benchmarks in Cybersecurity, Finance, and Law
• Evaluated via rigorous red-teaming with state authorities

API Pricing

Mistral AI has positioned Mistral Large 4 to be highly competitive, specifically targeting enterprise users who require high-throughput multimodal processing. The pricing structure is designed to reward efficient prompting through a significant discount on cached inputs, making long-context RAG (Retrieval-Augmented Generation) workloads much more affordable.

To celebrate the launch, Mistral has announced a 50% discount on all API usage for the first two weeks of the public preview. This provides an excellent opportunity for developers to benchmark the model against their current production workloads without the typical cost barrier.

• Input: $1.36 per 1M tokens
• Cached Input: $0.14 per 1M tokens
• Output: $4.18 per 1M tokens
• Launch Promo: 50% off for the first 14 days

Use Cases

The combination of a 1M context window and multimodal reasoning makes ML4 ideal for several high-impact domains. In software engineering, it can act as a sophisticated agent capable of performing vulnerability research, incident response, and complex refactoring across massive codebases. Mistral explicitly highlights that the eventual open-weight release will empower users to control their AI, which is critical for sensitive security research.

For legal and financial professionals, the model's ability to ingest and reason over massive document sets allows for automated due diligence, contract analysis, and complex regulatory compliance checking. In the realm of visual intelligence, the 1.6B vision encoder enables advanced applications in medical imaging analysis, industrial inspection, and automated video content auditing.

• Cybersecurity: Vulnerability research and incident response
• Legal & Finance: Large-scale document reasoning and compliance
• Software Engineering: Agentic coding and repository-wide refactoring
• Multimodal RAG: Complex visual and textual data retrieval

Getting Started

Developers can access the Mistral Large 4 public preview immediately through Mistral Studio. The model is available via a standard REST API, making it easy to integrate into existing Python or TypeScript workflows. For those looking to scale, the infrastructure is optimized for low-latency inference on Mistral's specialized European hardware.

We recommend starting with the Mistral AI SDK to explore the multimodal endpoints. As the model moves toward its full open-weight release at the end of October, developers should prepare their local deployment pipelines for a model of this scale.

• Access via Mistral Studio API
• Compatible with standard REST and Mistral SDKs
• Full weight release expected late October 2026

API Pricing — Input: $1.36 per million tokens / Output: $4.18 per million tokens / Context: 1M tokens

Sources