Mistral Large 3

Mistral did the thing the open-source community had been asking for since the Mixtral days: it put a flagship-class model under a license you can actually build a business on. Mistral Large 3, released December 2, 2025, is a 675-billion-parameter Mixture-of-Experts model under Apache 2.0. No research-only clause, no commercial license to buy, no asterisk. For a European lab that spent two years gating its best weights behind a restrictive license, that is the story. The model itself is good, not frontier, and we will be straight about which is which.

What changed: the license, mostly

Start with the part that matters to builders. Mistral's previous flagship, Mistral Large 2, shipped under the Mistral Research License: free to study, but you needed a paid commercial license to ship anything with it. Mistral Large 3 drops that entirely. The whole Mistral 3 family, including this 675B flagship, is Apache 2.0: use it commercially, self-host it, fine-tune it, redistribute it, no permission and no fee.

That puts Mistral back on the same open footing as DeepSeek, Qwen, and Llama, after a stretch where its openness story had quietly eroded. For anyone who wants a frontier-adjacent European model with clean commercial terms and EU data residency, this is the one.

What it actually is

Mistral Large 3 is a sparse Mixture-of-Experts model, Mistral's first big MoE since Mixtral. (An MoE routes each token to a few "expert" subnetworks, so you store all the parameters but only run a fraction per token.) The numbers: 675B total, about 41B active per token, which breaks down as a 673B language model plus a 2.5B vision encoder. It takes text and images, handles 40-plus languages, and carries a 256K-token context window. Mistral says it was trained from scratch on 3,000 NVIDIA H200 GPUs.

One thing it is not: a reasoning model. There is no long chain-of-thought mode baked in. That single fact explains most of the benchmark picture below.

Benchmarks: strong chat, not a leaderboard-topper

Be clear-eyed here. As a non-reasoning instruct model, Mistral Large 3 competes on general knowledge and chat, not on the hard reasoning and agentic-coding tests that reasoning-tuned models dominate. Mistral's own framing puts it at parity with the best instruction-tuned open models, second among open non-reasoning models on LMArena. Independent trackers are cooler.

Artificial Analysis Intelligence Index (independent). The index leans heavily on reasoning and agentic tasks, which Mistral Large 3 deliberately does not do, so it understates the model's value for plain chat and assistant work. Treat it as a reasoning-weighted ranking, not a verdict on general usefulness.

On GPQA-Diamond and SWE-Bench it trails the reasoning models badly, because those reward the step-by-step thinking it skips. It lands above Llama 4 Maverick and OLMo 3 on the same index. The takeaway: this is a competent, fast, multilingual assistant model, not the thing you reach for to solve a competition math problem or drive an autonomous coding agent.

Limitations and gotchas

• Not a reasoning model. If your task needs multi-step reasoning, agentic coding, or hard math, the reasoning-tuned open models beat it outright.
• 675B is not local. All experts must be resident; FP8 fits a single H200 or B200 node, NVFP4 a single H100 or A100 node. For your laptop, Mistral points to the smaller Ministral 3 family (14B, 8B, 3B).
• No GGUF or llama.cpp build at launch, and transformers support lagged. vLLM is the path; Ollama exposes it only as a cloud tag.
• Apache 2.0 governs the weights, though the model card adds a standard acceptable-use note.

Who should use it

Use it if you want a permissively licensed, multilingual, multimodal model for chat, drafting, extraction, and general assistant work, and you value EU provenance and clean commercial terms over leaderboard position. Reach for the API ($0.50 in, $1.50 out per million tokens) or a rented node; self-host only if you have the hardware. If you need reasoning or agentic coding, pair it with, or swap it for, a reasoning model like GLM-5.2 or DeepSeek V4. For local use, the smaller Mistral Small family is the better fit.

Run it in about 10 minutes

The fastest path is the hosted API. Self-hosting means a multi-GPU server.

# Quickest: the hosted API (la Plateforme). Set MISTRAL_API_KEY first.
curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"mistral-large-2512","messages":[{"role":"user","content":"Summarize this clause in plain English."}]}'

# Self-host FP8 on a single H200/B200 node with vLLM
vllm serve mistralai/Mistral-Large-3-675B-Instruct-2512 --tensor-parallel-size 8

If you only want to feel the licensing freedom, the more useful 10-minute move is to pull a small family member you can actually run, the Ministral 3 3B or Mistral Small, and confirm the Apache 2.0 license file is right there in the repo. That is the part you could not do with Large 2.

Sources and further reading

• Mistral 3 announcement (Mistral AI)
• Mistral Large 3 model card (Hugging Face)
• Mistral Large 3 docs
• The Register: Mistral 3 goes Apache 2.0
• Artificial Analysis: independent benchmarks

Tested on: not independently tested. Mistral Large 3 is a 675B MoE that needs a single H200/B200-class node even at FP8, beyond our bench. Mistral's own benchmark charts ship as images, so the comparison here uses independent Artificial Analysis figures, flagged as such. Sources linked above.
Date checked: 2026-06-26