Mistral Unveils Its Most Powerful Open Model Yet with Mistral 3

Mistral AI has launched Mistral 3, a new generation of open multimodal and multilingual models that includes three compact “Ministral” systems and its most advanced release yet, Mistral Large 3. All models are available under the Apache 2.0 licence.

The flagship Mistral Large 3 is a sparse mixture-of-experts model trained on 3,000 NVIDIA H200 GPUs, featuring 41 billion active parameters and 675 billion total parameters. 

The company claims that the model matches the performance of the strongest open-weight instruction-tuned systems and debuts as one among the top OSS non-reasoning models on the LMArena leaderboard.

The company said the model delivers improved image understanding and industry-leading multilingual performance outside English and Chinese. Both the base and instruction-tuned variants are available today, with a reasoning version expected soon.

The release is backed by a broader optimisation push involving NVIDIA, vLLM, and Red Hat. NVIDIA contributed Blackwell-optimised attention and MoE kernels, speculative decoding support, and low-precision TensorRT-LLM execution to ensure faster inference on Blackwell NVL72, A100, H100, and Jetson-class devices.

While Mistral Large 3 targets high-throughput enterprise workloads, the company is also releasing Ministral 3, a new edge-optimised family in 3B, 8B, and 14B sizes. Each variant comes in base, instruct, and reasoning formats, all with multimodal capabilities. 

Mistral claims the instruct versions often produce “an order of magnitude fewer tokens” than competing models, improving cost efficiency. The 14B reasoning model reaches 85% on AIME 2025, positioning it among the strongest mid-sized open models.

Mistral 3 is available on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, OpenRouter, and other platforms, with support for NVIDIA NIM and AWS SageMaker coming soon. Mistral also announced expanded custom-training services for enterprise deployments.

The company introduced Mistral Medium 3 earlier this year, a mid-sized model that surpasses rivals like Llama 4 Maverick while delivering over 90% of Claude Sonnet 3.7’s benchmark performance at a much lower cost. 

The model targeted enterprise workloads with stronger coding, STEM and multimodal capabilities, supports hybrid and on-premise deployments, and can run on as few as four GPUs. During this time, they hinted at a larger model in development, which was released as Mistral Large 3.

The post Mistral Unveils Its Most Powerful Open Model Yet with Mistral 3 appeared first on Analytics India Magazine.