Mistral Large 3

Summary

Mistral Large 3 is Mistral AI's Apache 2.0 open-weight flagship, released December 2, 2025 with a 41B-active / 675B-total granular MoE architecture and 256K context. At $0.50/$1.50 per million tokens, it's among the most cost-effective frontier-class open models from any major lab.

Overview

Mistral Large 3 is Mistral AI's flagship open-weight frontier model, released December 2, 2025 under the Apache 2.0 license. With 41 billion active parameters across a "granular Mixture of Experts" architecture (675 billion total parameters), it is the most capable model in the Mistral 3 family and the first open-weight model to offer genuine frontier-level performance with full multimodal and multilingual capability in a single model. At $0.50/$1.50 per million tokens, it is also one of the most cost-effective frontier-class models available from any major lab.

Mistral Large 3 is Europe's strongest answer to the US-dominated frontier AI landscape. Its Apache 2.0 license means it can be used commercially without restrictions, fine-tuned freely, and deployed on-premises — advantages that distinguish it sharply from proprietary models.

Specifications

  • Developer: Mistral AI (Paris, France)
  • Model String: mistral-large-latest / mistral-3-large (check Mistral docs)
  • Release Date: December 2, 2025
  • Type: Multimodal LLM, Mixture-of-Experts, Open-Weight
  • Architecture: Granular MoE — 41B active parameters / 675B total parameters
  • Context Window: 256,000 tokens
  • Modalities: Text and image input; text output
  • Languages: Multilingual (40+ languages)
  • License: Apache 2.0 (fully open, commercial use permitted)
  • Access: Mistral API (la Plateforme), Hugging Face, Azure AI Foundry, AWS Bedrock, Google Cloud, self-hosted
  • Pricing: $0.50 per million input tokens / $1.50 per million output tokens

Capabilities

Open-Weight Frontier Performance: Matches important capabilities of GPT-4o and Gemini 2.0 Flash at launch — competitive with closed models from US labs while being fully open source and self-hostable.

Multimodal: First Mistral model with native vision capability — handles image understanding, document analysis, and visual reasoning alongside text.

Multilingual: Strong performance across 40+ languages, positioning it as particularly valuable for non-English deployments where US models may underperform.

Granular MoE Architecture: The "granular" MoE design provides more fine-grained expert routing than standard MoE approaches, contributing to both efficiency and quality.

Cost Efficiency: At $0.50/$1.50, one of the best price/performance ratios for a frontier-class model. Combined with Apache 2.0 licensing, total cost of ownership for self-hosted deployments is extremely low.

Code: Complemented by Codestral for coding-specific tasks, but Mistral Large 3 itself has strong general coding capability.

Limitations

Benchmark performance against the Gemini 3 and Claude 4-series models (released after Mistral Large 3) shows a gap — Mistral has not published direct comparisons against these newer models. The 256K context window, while generous, is smaller than the 1M+ options now standard among Anthropic, OpenAI, and Google flagship models.

Recent Developments

  • December 2, 2025 Launch: Part of the broader Mistral 3 release alongside nine Ministral 3 small models — a comprehensive portfolio refresh.
  • Apache 2.0 License: The choice of fully open licensing was deliberate positioning against US labs, reinforcing Mistral's identity as the open-source European alternative.
  • €1.7B Series C (September 2025): Raised at €11.7B valuation led by ASML, providing significant runway for continued model development and infrastructure.
  • €1.2B Sweden Data Center: Investing in European AI infrastructure to support sovereign deployments.

Last Updated

February 26, 2026

→ Back to Models