Mistral 3 Review: Large 3 MoE + Ministral Edge Models (December 2025)

Published December 2, 2025 — Updated June 11, 2026

On December 2, 2025, Mistral AI released Mistral 3 — a new model family consisting of Mistral Large 3, their first sparse mixture-of-experts (MoE) model since the Mixtral series, and the Ministral 3 series of compact edge models (3B, 8B, 14B). Every model ships under the Apache 2.0 license. Here’s everything you need to know.

What’s in Mistral 3?

Mistral 3 is not a single model — it’s a two-tier release. At the top sits Mistral Large 3, a sparse MoE with 675B total parameters and 41B active parameters per token. Below it, the Ministral 3 family offers dense models at 3B, 8B, and 14B parameters, each available in base, instruct, and reasoning configurations.

This structure mirrors what Mistral did with the Mixtral series: a large expert model at the frontier and smaller dense models for latency-sensitive and on-device scenarios. The key difference is that all Mistral 3 models include multimodal image understanding across 40+ languages.

Mistral Large 3: The MoE Flagship

Mistral Large 3 is Mistral’s most capable model. It uses a sparse mixture-of-experts architecture with 675B total parameters and 41B active parameters per inference step — meaning inference cost is much lower than a dense 675B model, while quality approaches that scale.

It ranks #2 on LMArena among OSS non-reasoning models, making it competitive with the strongest open-weight alternatives in real-world chat quality. Mistral describes it as achieving “parity with the best instruction-tuned open-weight models on the market on general prompts.”

Like the Ministral models, Mistral Large 3 supports multimodal input. The context window was not disclosed in the launch announcement; check the Mistral API docs for current limits.

Ministral 3 Series: Edge-First

The Ministral 3 family covers three weight classes: 3B for embedded and on-device inference, 8B for consumer GPUs and local servers, and 14B for edge servers requiring higher capability. Each size ships as base, instruct, and reasoning variants.

The standout result is the Ministral 14B reasoning variant at 85% on AIME ’25 — a challenging formal mathematics benchmark. That score puts a 14B model in competition with much larger models, making it the go-to option for math reasoning, code analysis, and formal logic tasks at edge scale.

Benchmarks at a Glance

Apache 2.0: Why It Matters

Mistral has released all Mistral 3 models under the Apache 2.0 license — one of the most permissive open licenses available. This means you can download the weights, fine-tune, deploy commercially, and redistribute without restriction.

This is a significant commercial advantage over models like Llama (which uses a custom Meta license restricting redistribution above certain usage thresholds) or Mistral’s own Medium 3.5 (modified MIT). For enterprises that need a legally clean open-weight frontier model, Mistral 3 is currently the strongest option available.

Pricing

All Mistral 3 weights are free to download and self-host under Apache 2.0. API access is available via the Mistral API with pay-per-token pricing — check mistral.ai/technology/#pricing for current rates, as Mistral updates these regularly. The models are also accessible through Le Chat on free and Pro plans.

For self-hosted deployments: Mistral Large 3 requires substantial multi-GPU infrastructure. The Ministral 3B and 8B models can run on a single consumer GPU (or quantized on CPU), while the 14B targets single high-end GPU setups.

Who Should Use Mistral 3?

• Open-weight advocates: Apache 2.0 makes Mistral Large 3 the most commercially flexible frontier-class MoE available.
• Edge/on-device teams: Ministral 3B and 8B are purpose-built for constrained hardware with no compromise on language breadth (40+ languages).
• Math and reasoning apps: Ministral 14B reasoning at 85% AIME ’25 is the strongest sub-20B math reasoning model in the open-weight ecosystem.
• Multilingual products: Image understanding across 40+ languages in a compact model is rare — Ministral covers that gap.

vs. Mistral Small 4 and Medium 3.5

Mistral’s lineup has expanded rapidly. Mistral 3 (December 2025) was followed by Mistral Small 4 (March 2026) and Mistral Medium 3.5 (May 2026). Small 4 unified reasoning, vision, and coding into a single 119B MoE model at lower cost; Medium 3.5 is a dense 128B merged model focused on agentic coding (77.6% SWE-Bench). Mistral 3 remains the right choice when you need Apache 2.0 licensing or compact sub-15B edge deployment.

If you’re choosing between them for API use, compare current pricing on the Mistral API docs — the per-token rates differ by tier.

Frequently Asked Questions