Now loading...
Mistral AI has unveiled Mistral 3, its latest lineup of advanced artificial intelligence models designed to advance open-source capabilities. The series features three compact dense models with 14 billion, 8 billion, and 3 billion parameters, alongside the flagship Mistral Large 3, a sparse mixture-of-experts system boasting 41 billion active parameters and 675 billion total. All variants are made available under the permissive Apache 2.0 license, with compressed formats to facilitate widespread adoption among developers and promote accessible distributed computing.
The smaller Ministral models stand out for delivering superior performance relative to their computational costs, positioning them as leaders in efficiency for their size category. Meanwhile, Mistral Large 3 enters the competitive field of top-tier open-source instruction-tuned models, marking a significant evolution in the company’s pretraining techniques.
Mistral Large 3 ranks among the world’s leading openly available large language models, having been developed from the ground up using 3,000 NVIDIA H200 graphics processing units. As the firm’s first mixture-of-experts architecture following the influential Mixtral family, it matches the capabilities of premier open-weight models in handling general queries. The system also excels in processing images and leads in multilingual dialogues beyond English and Chinese contexts. Following refinement, it secures second place in the open-source non-reasoning category and sixth overall on the LMSYS Arena leaderboard.
Both foundational and instruction-adjusted editions of Mistral Large 3 are now public, offering robust bases for enterprise adaptations and developer innovations. An enhanced reasoning edition is slated for imminent release.
To improve accessibility, Mistral AI has collaborated with vLLM, Red Hat, and NVIDIA to optimize deployment. A specialized checkpoint in NVFP4 format, created using the llm-compressor tool from vLLM’s repository, enables smooth operation of Mistral Large 3 on Blackwell NVL72 setups or single clusters of eight A100 or H100 GPUs via vLLM.
This partnership with NVIDIA extends to training the entire Mistral 3 family on Hopper GPUs, using high-bandwidth HBM3e memory for demanding tasks. NVIDIA’s integrated design philosophy unites hardware, software, and models, with support added for efficient low-precision inference through TensorRT-LLM and SGLang. For the MoE structure of Large 3, optimizations include Blackwell attention mechanisms, disaggregated serving for prefill and decode phases, and joint work on speculative decoding to handle extended contexts and high-volume processing on GB200 NVL72 platforms. On the device side, Ministral models are tuned for NVIDIA’s DGX Spark workstations, RTX-powered personal computers and laptops, and Jetson embedded systems, ensuring consistent performance across environments from data centers to autonomous devices.
Mistral AI expressed gratitude to these partners for their contributions in making the models more efficient and reachable.
Tailored for on-device and localized applications, the Ministral 3 lineup includes 3 billion, 8 billion, and 14 billion parameter options, each offered in base, instruction-tuned, and reasoning configurations. All incorporate image comprehension and operate under Apache 2.0, combining native support for multiple modalities and over 40 languages to suit diverse business and creative demands.
These models achieve the highest efficiency in open-source terms, balancing output quality with reduced token generation and smaller footprints compared to peers. In precision-focused scenarios, the reasoning versions extend their deliberation for top accuracy within their scale, such as the 14 billion parameter model reaching 85 percent on the AIME 2025 benchmark.
Mistral 3 models are immediately accessible through Mistral AI Studio, Amazon Bedrock, Azure AI Foundry, Hugging Face collections for Large 3 and Ministral, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI. Upcoming integrations include NVIDIA NIM and AWS SageMaker.
For businesses needing bespoke AI, Mistral AI provides custom training to refine models for specific domains, datasets, or infrastructures, ensuring secure, scalable implementations. Developers can begin with documentation on the AI Governance Hub, experiment via Hugging Face or Mistral’s platform with API access, pursue customizations through direct contact, and engage the community on Twitter, Discord, or GitHub.
This release underscores Mistral AI’s commitment to open innovation, enabling frontier-level intelligence, cross-modal applications, and adaptable scaling from compact devices to large-scale operations.
