Mistral 3 Unveiled: How Its New Open‑Source Models Redefine Performance and Cost

Recently, Mistral AI open‑sourced Mistral 3, the next generation of its models.

Mistral 3 includes three industry‑leading small dense models (14B, 8B, 3B) and the most powerful Mistral Large 3, a sparse mixture‑of‑experts (MoE) model with 41 B active parameters and 675 B total parameters.

Benchmark comparisons are made against two domestic models, DeepSeek‑3.1 (670 B) and Kimi‑K2 (1.2 T), where Mistral 3 leads across multiple dimensions.

Mistral Large 3: A New Open‑Source Benchmark

Mistral Large 3 is among the strongest permissive open‑source weight models, trained from scratch on 3,000 NVIDIA H200 GPUs. Building on the pioneering Mixtral series, this MoE architecture marks a major leap in pre‑training. After instruction fine‑tuning, it matches top instruction‑tuned open‑source models on general prompts, supports image understanding, and excels in multilingual (non‑English/Chinese) dialogue.

On the LMArena leaderboard, Mistral Large 3 ranks #2 among non‑inference open‑source models and #6 overall.

Ministral 3: Edge‑Optimized Intelligence

Targeting edge and on‑premise scenarios, the Ministral 3 series offers 3 B, 8 B, and 14 B models, each with base, instruct, and reasoning variants, all supporting image understanding and released under the Apache 2.0 license. Combined native multimodal and multilingual abilities enable coverage of any enterprise or developer need.

The key advantage is Ministral 3’s lowest cost‑performance ratio among open‑source models. In real deployments, token generation cost scales with model size; the Ministral instruct variant achieves comparable performance while generating an order of magnitude fewer tokens.

https://huggingface.co/collections/mistralai/mistral-large-3
https://mistral.ai/news/mistral-3