Mistral AI Unveiled Mistral 3: New Series of Models with Mistral Large 3

Mistral AI unveiled Mistral 3 — a new series of open-source language models in various sizes. The family includes compact Ministral 3 models (3B, 8B, 14B) for local use and a powerful Mistral Large 3 (675B parameters) for complex tasks. All models are released under Apache 2.0 and support text, images, and multilingual queries.

Mistral Large 3: A New Frontier

Mistral Large 3 is the flagship of the series, trained from scratch on 3,000 NVIDIA H200 GPUs. This is Mistral's first model using sparse mixture-of-experts (MoE) architecture with 41B active parameters out of 675B total. On LMArena benchmarks, Mistral Large 3 ranks second among open language models and demonstrates results comparable to the best instruction-tuned models on the market.

The key advantage of MoE architecture is that the model doesn't use all parameters simultaneously. Instead, different parts of the network activate for different types of queries — this makes inference faster and cheaper than fully utilizing 675B parameters. This approach allows scaling models without proportional increases in computational resource requirements.

The model demonstrates particularly strong results in multilingual tasks and image understanding. The company promises to soon release a version with enhanced logical reasoning capabilities and deeper analysis of complex problems.

Partnership with NVIDIA for Speed and Scaling

Mistral worked with NVIDIA, vLLM, and Red Hat to optimize inference and model deployment. All Mistral 3 models were trained on NVIDIA Hopper GPUs, which enabled the use of high-bandwidth HBM3e memory — a critical component for working with such massive neural networks. NVIDIA created specialized optimizations in TensorRT-LLM and SGLang for efficient instruction execution.

For Mistral Large 3, support for efficient Blackwell cores was added and the attention/MoE architecture was improved for long contexts on GB200 NVL72 systems. This enables serving high-performance workloads with minimal latency.

Compact Ministral models can be easily deployed on local machines:

• On DGX Spark for enterprise solutions
• On RTX PCs and laptops for development
• On Jetson devices for IoT and robotics
• Support for deployment from cloud infrastructure to edge devices

This vertical integration means developers get a unified path for running the same models from data centers to local edge devices without rewriting code.

Ministral 3: A Powerful Tool for Edge

For edge computing and local use, Mistral released Ministral 3 in three sizes: 3B, 8B, and 14B parameters. Each size is available in three variants: a base model, an instruction-tuned version for following instructions, and a version with enhanced logical reasoning capabilities.

All variants support images and text in 30+ languages, including Turkic languages and Russian. Despite its compact size, Ministral 3 delivers industry-leading performance-to-cost ratio among open models. This is critical for companies that want to run AI locally without cloud services.

What This Means for the AI Industry

Open language models are becoming more practical and accessible. With an Apache 2.0 license, anyone can use, modify, and develop Mistral models in commercial projects without restrictions. For developers, this means more flexibility in stack selection; for enterprises, it means reduced AI infrastructure costs and less dependence on cloud providers. Mistral 3 could be a turning point in the move toward independent, locally-managed AI systems.