Mistral Large 3

Mistral Large 3, released December 2 2025 as part of the Mistral 3 family, is Mistral AI's flagship sparse mixture-of-experts (MoE) model with 675B total parameters but only 41B active per token. It delivers frontier-level performance across general knowledge, coding, mathematics, and multilingual tasks while maintaining exceptional cost-efficiency — it can run on a single 8×GPU node via NVFP4 quantization. All models in the Mistral 3 family are released under the Apache 2.0 license, making Large 3 the most capable fully open-weight frontier model at the time of its launch.

From a TheRouter operator perspective, Large 3 sits alongside Mistral's smaller Ministral 3 models and its reasoning-enhanced Magistral family. Its sparse MoE architecture (41B active / 675B total) makes it a strong default for open-weight self-hosting scenarios where per-token efficiency matters more than peak single-query reasoning depth. For tasks requiring the hardest multi-step reasoning, Mistral's own Magistral Small (24B, reasoning-enhanced) or Anthropic/OpenAI frontier models may be more appropriate — Large 3 scores ~43.9% on GPQA Diamond vs Claude's 65% — but for high-volume production workloads, large document processing, and multilingual agent applications, Large 3's combination of open weights, 256K context, and $0.50/$1.50 pricing through TheRouter is hard to beat.