Ministral 3 14B

Ministral 3 14B is the top of the three-model Ministral 3 line that Mistral AI published on 2 December 2025 alongside its flagship Mistral Large 3. Where Large 3 is a 675-billion-parameter mixture-of-experts for server rooms, the Ministral models are dense networks sized to fit on hardware a single developer owns: laptops with discrete GPUs, NVIDIA Jetson boards, DGX Spark desktops. The 14B variant carries 13.9 billion parameters including its vision encoder, accepts text and images, and holds a 256,000-token context — the same window as the lab's much larger models. Mistral publishes each size in three flavours under one Apache 2.0 licence: a base model for further training, an instruct model for chat and tool use, and a reasoning model that thinks before answering. The reasoning variant is where the 14B earns its place; the lab reports 85% on AIME 2025, a competition-mathematics benchmark, which is a result normally associated with far heavier models. The vendor positions the instruct version as comparable to its own Mistral Small 3.2 at 24 billion parameters, which is the practical claim: roughly the previous generation's mid-size quality at 40% fewer weights. Hosted access costs 0.20 US dollars per million tokens, identical for input and output — Mistral prices the whole Ministral line symmetrically, which is unusual and makes long generations cheap relative to competitors that charge three to five times more for output. The technical report describing the family arrived later, on 13 January 2026, as arXiv 2601.08584.

▸Videos

No videos yet.

▸Reviews