Launch of Mistral 3 Series

Mistral AI has released the Mistral 3 series of open-source models along with Mistral Large 3. - Mistral Large 3 features a MoE architecture with a total of 675 billion parameters and 41 billion active parameters. - Large 3 supports image input, multimodal dialogue, and native support for over a dozen major languages. - Large 3 accommodates a context length of 256K tokens, native function calls, and structured JSON output. - The Mistral 3 series includes small dense models of 3B, 8B, and 14B, designed for edge deployment and cost-effectiveness. - The 14B inference version achieved an accuracy of 85% in AIME '25 mathematics competition-level tests. - Mistral has collaborated with NVIDIA, Red Hat, and vLLM to release optimized checkpoints to reduce deployment costs. - The official statement claims these models lead in open-source multilingual and multimodal capabilities while optimizing single-node operation. - According to the LMArena leaderboard, Mistral Large 3 ranks 2nd among open-source non-inference models and 6th overall in open-source. - Edge memory requirements: 14B requires approximately 24GB, 8B requires approximately 12GB, and 3B requires approximately 8GB.
Key insights
Mistral Covers Edge and Cloud Scenarios with a Multi-Tier Product Line
Mistral AI has simultaneously launched Mistral Large 3 (675B MoE) and the smaller Mistral 3 models (3B/8B/14B), creating a product line that spans from edge to cloud. The Mistral 3 series is positioned for high cost-effectiveness and edge deployment, with 3B/8B/14B models requiring 8GB/12GB/24GB of memory respectively. Large 3, being a sparse expert mixture architecture, targets workflows that require ultra-long context and complex tool calls, forming a complementary array with the smaller models.
Technical Specifications and Multimodal Capabilities of Mistral Large 3
Mistral Large 3 utilizes a MoE architecture with a total of 675 billion parameters and 41 billion active parameters. It natively supports image input, multimodal dialogue, and over a dozen major languages, while also accommodating a context length of 256K tokens, native function calls, and structured JSON output for driving external systems and complex workflows.
Outstanding Performance of Small Model Inference Versions Reduces Cloud Dependency
The Mistral 3 small models have released dedicated inference versions, with the 14B inference version achieving an accuracy of 85% in AIME '25 mathematics competition-level tests. This indicates significant progress in certain deep inference tasks, allowing deployment on edge devices such as laptops and robots, thereby reducing reliance on expensive cloud-based large models.
Ecosystem and Deployment Optimization Achieved Through Industry Collaboration
Mistral has engaged in deep adaptation with infrastructure vendors such as NVIDIA, Red Hat, and vLLM, releasing optimized checkpoints and single-node operation strategies to enhance inference efficiency and simplify the deployment process of large models with fewer hardware resources.
Benchmarks and Rankings Demonstrate Open-Source Competitiveness
In the LMArena leaderboard, Mistral Large 3 ranks 2nd among open-source non-inference models and 6th overall in open-source. Additionally, reports indicate its performance surpasses that of Kim 2, showcasing its leading position within the open-source community.