AI startup Mistral has unveiled Mistral 3, the latest generation of its models, featuring three compact, high-performance dense models of 14B, 8B, and 3B parameters, alongside Mistral Large 3, its most advanced model to date—a sparse mixture-of-experts system trained with 41B active and 675B total parameters. All models are available under the Apache 2.0 license, providing developers with open-source access in multiple compressed formats to support distributed AI applications.
The Ministral models are designed for strong performance-to-cost efficiency, while Mistral Large 3 positions itself among leading instruction-fine-tuned open-source models. Trained from scratch on 3,000 NVIDIA H200 GPUs, Mistral Large 3 marks the company’s first mixture-of-experts release since the Mixtral series and represents a significant advancement in pretraining. After post-training, it matches top instruction-tuned open-weight models on general prompts and demonstrates advanced image understanding as well as superior multilingual conversation capabilities.
Mistral Large 3 debuted at #2 in the OSS non-reasoning models category and #6 overall on the LMArena leaderboard. Both base and instruction-tuned versions are released under Apache 2.0, offering a robust platform for enterprise and developer customization, with a reasoning version planned for future release.
Mistral Large 3 has been made highly accessible to the open-source community through collaborations with vLLM and Red Hat. A checkpoint in NVFP4 format, optimized with llm-compressor, allows efficient execution on Blackwell NVL72 systems or a single 8×A100 or 8×H100 node using vLLM.
The development of advanced open-source AI models relies on extensive hardware-software optimization, achieved in partnership with NVIDIA. All Mistral 3 models, including Large 3 and Ministral 3, were trained on NVIDIA Hopper GPUs, utilizing high-bandwidth HBM3e memory for large-scale workloads. NVIDIA’s co-design approach integrates hardware, software, and models to enable efficient inference using TensorRT-LLM and SGLang across the Mistral 3 family, supporting low-precision execution.
For the sparse mixture-of-experts architecture of Large 3, NVIDIA implemented Blackwell attention and MoE kernels, added prefill/decode disaggregated serving, and collaborated on speculative decoding, enabling developers to handle long-context, high-throughput workloads on GB200 NVL72 systems and beyond. Ministral models are also optimized for deployment on DGX Spark, RTX PCs and laptops, and Jetson devices, providing a consistent, high-performance experience from data centers to edge applications. Mistral extends thanks to vLLM, Red Hat, and NVIDIA for their support and collaboration.
The Ministral 3 series is designed for edge and local deployments, offered in three sizes—3B, 8B, and 14B parameters. Each size is available in base, instruct, and reasoning variants, all featuring image understanding and released under the Apache 2.0 license. Combined with native multimodal and multilingual capabilities, the Ministral 3 family provides versatile solutions for both enterprise and developer applications.
The series delivers an exceptional cost-to-performance ratio among open-source models, with instruct variants matching or surpassing comparable models while generating significantly fewer tokens. For scenarios where accuracy is paramount, the reasoning variants can perform extended computations to achieve leading accuracy within their weight class, such as 85% on AIME ’25 with the 14B model.
Mistral 3 is currently accessible through Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face (Large 3 & Ministral), Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI, with availability on NVIDIA NIM and AWS SageMaker coming soon.
Mistral remains a leading contributor to Europe’s AI model ecosystem and open-source initiatives, though its latest flagship model still lags behind top industry competitors in terms of performance, speed, and cost. The smaller Ministral variants may offer a more practical alternative, providing flexible options for diverse use cases and deployment across different devices.
The post Mistral Unveils Mistral 3: Next-Gen Open-Source Models For Enterprise And Edge AI appeared first on Metaverse Post.
