French AI startup Mistral has released its most ambitious model family to date, positioning itself as a formidable open-source competitor against industry leaders like DeepSeek. The release, announced on Tuesday, consists of four models under the permissive Apache 2.0 license, ranging from a compact 3-billion parameter version to the flagship Mistral Large 3 with 675 billion parameters.
The flagship model employs a sparse Mixture-of-Experts architecture, activating only 41 billion of its total parameters per token. This design allows it to compete with frontier models while maintaining an inference compute profile closer to a 40-billion parameter model. Mistral Large 3 was trained from scratch on 3,000 NVIDIA H200 GPUs and debuted at number two among open-source, non-reasoning models on the LMArena leaderboard.
Benchmark comparisons with DeepSeek reveal a nuanced rivalry. According to Mistral's data, its best model outperforms DeepSeek V3.1 on several metrics but slightly trails the newer V3.2 on the LMArena. While DeepSeek maintains an edge in raw coding speed and mathematical logic, the Mistral family holds its own in general knowledge and expert reasoning tasks.
The smaller "Ministral" models (3B, 8B, and 14B parameters) offer significant potential for developers. The 3B model, noted by AI researcher Simon Willison, can run entirely in a browser via WebGPU, opening possibilities for drones, robots, offline laptops, and embedded systems. Early testing shows the Mistral 3 Large excels in conversational fluency with a natural cadence and minimal censorship, making it suitable for creative writing and role-play.
Enterprise adoption is already underway, with HSBC announcing a multi-year partnership with Mistral to deploy generative AI across its operations. The bank will run self-hosted models on its own infrastructure, leveraging Mistral's expertise while maintaining data control under GDPR compliance. Mistral also collaborated with NVIDIA on an NVFP4 compressed checkpoint, enabling Large 3 to run on a single node of eight high-end GPUs. NVIDIA claims the Ministral 3B achieves roughly 385 tokens per second on an RTX 5090.
A reasoning-optimized version of Large 3 is expected soon. For enterprises seeking frontier capability with open weights, multilingual strength across European languages, and a vendor not subject to Chinese or American national security laws, Mistral now represents a viable option.
