Mistral Large 3: Europe's frontier model with open weights
Mistral Large 3 has been available since December 2, 2025 — Apache 2.0 license, 675 billion total parameters, 41 billion active per forward pass. With it, Mistral is the first European vendor to ship a model at frontier level whose weights aren’t only reachable through an API, but downloadable and commercially usable. The model identifier mistral-large-2512 puts the training date right into the name — semi-annual cycle, not monthly.
What makes Large 3 distinct
- Sparse Mixture-of-Experts with 675B total / 41B active parameters. Trained from scratch on 3,000 NVIDIA H200 GPUs.
- Apache 2.0 license for both base and instruction-tuned variants. Commercial use, fine-tuning and self-hosting without royalty.
- Multimodal from day one — every variant handles text and images, natively trained on 40+ languages. Context window: 256K tokens.
- #2 in the OSS non-reasoning category on LMArena, #6 across all open-weight models. MMLU-Pro: 73.11%, MATH-500: 93.60%.
What used to be
Through the end of 2025, the picture was clear: the genuinely strong frontier models were proprietary (GPT-4, Claude 3, Gemini 1.5), and the genuinely open ones (Llama 3, Qwen 2.5, Mistral Large 2) sat one or two steps below. If you wanted “GPT-4 level” and “Apache 2.0” at the same time, you simply couldn’t have it.
Mistral Large 2 from July 2024 was open, but under the “Mistral Research License” — commercial use required a separate contract. In practice, that was a sticking point for companies that wanted to self-host without licensing negotiation.
What now applies
1. True frontier level under Apache 2.0. Mistral Large 3 is one of the first open-weight models that clearly moves to the front of the OSS field on the LMArena leaderboard — under a license that’s permissive enough for any practical use. Anyone who needs their own servers, their own clusters, or air-gapped environments can run it without vendor lock-in. Mistral’s site reports MMLU-Pro at 73.11% and MATH-500 at 93.60%.
2. MoE as an efficiency bet. At 41B active parameters, Large 3 lands closer to a 50B dense model on inference cost than to a classic 70B+. Memory footprint (all 675B parameters need to be loaded) stays high though — a single H100/H200 won’t do it; in practice you need a multi-GPU setup. If you can afford the hardware, you get more quality per token at comparable latency.
3. Multimodal and multilingual by default. Unlike earlier open-weight generations, image understanding and 40+ languages aren’t separate models — they’re part of the main one. Mistral explicitly highlights that Large 3 is particularly strong in non-English, non-Chinese languages, which is relevant for European workloads where German, French, Spanish or Italian matter more than the English benchmark average.
Context
Strategically, this release fits Mistral’s positioning as a sovereign-AI player. Anyone who not only consumes but actually runs a European model at Apache 2.0 level builds a real alternative to API dependence on US vendors. For regulated industries (finance, health, public sector), that argument outweighs two points on an eval score.
What the hype skips: frontier performance on your own hardware is expensive. Even moderate inference load on Large 3 demands 8x H100 or equivalent — investments that, for most mid-sized companies, run cheaper through a European cloud host (Scaleway, OVH, IONOS). So the choice isn’t “API vs. own hardware” — it’s “US API vs. EU API on the same model weights.”
What you can do now
If you want to seriously test open weights: Pull the weights from Hugging Face and try them on your top-3 tasks. Key question: is the latency in a self-hosted setup acceptable, or is an API through a European host the more pragmatic choice?
If you need GDPR compliance or data residency: Mistral Large 3 via a European inference provider is currently the best combination of quality and data sovereignty. Apache 2.0 makes vendor lock-in nearly impossible.
If you just want one API vendor: Test Large 3 as a default for multilingual, multimodal workloads — and evaluate it against GPT-5 or Claude Opus 4.7. On English benchmarks the US models look ahead; on everyday German prose, they sit closer together than benchmarks suggest.
Entdecke mehr
Three Days with Claude Fable 5 — and Why Opus Stays My Daily Driver
Fable 5 fixed a bug in one shot that Opus failed at twice. I'm still not switching. My honest impressions from the trenches after three days.
GlossarMistral Large 3
Mistral Large 3 (December 2025) is Mistral's open-weight MoE flagship — 675B parameters, 256K-token context, natively multimodal, Apache 2.0.
LexikonComparing AI models — who builds what and how to choose
The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.