Ministral 3 14B

Ministral 3 14B is the largest of Mistral AI's Ministral 3 family, released December 2025 in Base, Instruct, and Reasoning variants. It is a dense, edge-optimized model pairing a language backbone with a 0.4B-parameter vision encoder for multimodal understanding, with up to 256K token context via YaRN RoPE scaling.

Mistral positioned the family as models small enough to run on drones, laptops, and edge hardware, with three sizes (3B, 8B, 14B) and Apache 2.0 open weights. The 14B reasoning variant is reported by trackers to reach 85 percent on AIME 2025, a striking result for an edge-class model. It is served by vLLM and supported for fine-tuning via standard tooling.