Overview
- Tool name: Mistral Large 3
- Developer: Mistral AI (France)
- Official website: https://mistral.ai
- Category: AI Models
- License: Apache 2.0 (open weights)
Mistral Large 3 is the flagship model from European lab Mistral AI: a 675B-parameter sparse mixture-of-experts (MoE) model with 41B parameters active per token, a 256K-token context window, image input, tool calling and structured output. It also marks the first time a European frontier lab reopened weights at the flagship tier - the Apache 2.0 license permits commercial use, fine-tuning and self-hosting, which matters for enterprises (banking, healthcare, public sector) concerned with data residency and vendor lock-in.
Officially reported benchmarks: MMLU-Pro 84.1, GPQA Diamond 74.6, SWE-bench Verified 62.3, AIME 2025 88.0 (with reasoning traces). Aggregate capability sits a few points behind GPT-5.5 and Claude 4.7, but at the top of the open-weights class.
The wider Mistral lineup includes Medium 3.5 (long-horizon agentic coding, Modified MIT, $1.5/$7.5), Small 4 (Apache 2.0, $0.15/$0.6), the Ministral 3 series (3B/8B/14B for the edge), Codestral (low-latency coding) and Voxtral (speech).
API pricing (official, per million tokens)
- Input $0.50 / output $1.50, cached input $0.05
- Batch at 50% off; regional inference at 1.1x list
- Roughly 65% of GPT-5.5's cost at comparable capability
Who it is for
- Enterprises needing self-hosted frontier capability: weights are downloadable and run on private infrastructure
- European compliance scenarios: EU data residency and EU AI Act documentation
- High-throughput production: 90% discount on cached input keeps costs manageable
Limitations
- Still a few points behind the top closed flagships on the hardest reasoning tasks
- Self-hosting a 675B model is demanding: quantization and multi-GPU setups are the norm, and self-hosting usually only pays off above tens of billions of tokens per month
- Benchmark figures are vendor-reported; third-party SWE-bench replications run lower (around 58)
Frequently asked questions
Is Mistral Large 3 open source?
Open weights under Apache 2.0 - commercial use, fine-tuning and self-hosting allowed, with no revenue-share clause (unlike Llama's community license).
What does the API cost?
Official list price is $0.50 per million input and $1.50 per million output tokens, cached input at $0.05, batch at half price.
Mistral Large 3 or Medium 3.5?
Large 3 is the multimodal flagship (with vision); Medium 3.5 targets long-horizon agentic and coding work (Modified MIT, $1.5/$7.5). Pure-text agent workflows are often well served by Medium 3.5.
*This page is not affiliated with Mistral AI. Specs and pricing verified against mistral.ai as of October 2026; confirm current figures on the official site.*
