Mistral Large 3: Apache 2.0 Open Weights With a 256k Context

What is Mistral Large 3?

Mistral Large 3 is an open-weight general-purpose multimodal model from the French company Mistral AI, published as version 25.12 on 2 December 2025 under the Apache 2.0 licence. It uses a mixture-of-experts design totalling 675 billion parameters with 41 billion active, and takes a 256,000 token context window. Because the weights and licence are both public, it is one of the two models on this site you can genuinely download and run yourself - the other being Llama 4 Scout.

Efficiency by Design

Mistral built its reputation on getting more from less, and Large 3 embodies that philosophy. Using an efficient mixture-of-experts architecture, it activates only a portion of its parameters per token, achieving high throughput and low latency. For high-volume, latency-sensitive production workloads - where every millisecond and every dollar of inference cost matters - this efficiency is a decisive advantage over heavier models of similar quality.

What Mistral's Model Card States

Mistral's own model card for v25.12 documents the licence, parameter counts and context window used throughout this guide. It does not publish function-calling benchmark scores or multilingual performance figures, so this guide does not claim strength in either - an earlier version asserted both without a source.

Licensing

The Apache 2.0 licence is a standard, well-understood open-source licence: it permits commercial use, modification and self-hosting with minimal restriction, and does not require derivative works to be open-sourced in turn. That much is a fact about the licence itself, not a claim about Mistral or this model specifically.

Deployment and Cost

Being Apache 2.0 with published weights, Mistral Large 3 can be self-hosted at infrastructure cost alone, or accessed through Mistral's own API for teams that prefer managed hosting. Beyond that, this guide does not claim specific inference-stack support or hosting-provider availability without checking each one directly.

Ideal Use Cases

Mistral Large 3 fits scenarios where open weights and self-hosting matter - data sovereignty requirements, privacy-sensitive deployments, or wanting to fine-tune on private data. Against the other open-weight option here, Llama 4 Scout, it trades a much smaller context window (256k vs 10M) for a smaller, more manageable model footprint.

Mistral Large 3 vs Competitors

Of the two genuinely self-hostable models on this site, Mistral Large 3 has the far smaller context window (256k against Llama 4 Scout's 10M) but is the only one of the two with a strong LMArena entry: 1414, rank 125 as of 23 August 2026 - well below the closed frontier tier, though still a real, usable open-weight option depending on what you need it for.

Official Resources