Ministral 3 3B

Ministral 3 3B is the entry point to the Ministral 3 family released by Mistral AI on 2 December 2025, and the most used of the three. In the 30 days to 3 September 2026 its instruct version was downloaded roughly 352,000 times from Hugging Face, ahead of both the 14B (212,000) and the 8B (138,000) — a reminder that in open weights the deciding factor is rarely the benchmark table but whether the file fits on the machine at hand. What is notable is how little the size costs in features. At 3.8 billion parameters the model still reads images, still holds a 256,000-token context, still calls functions and returns JSON, and still comes in base, instruct and reasoning variants under a single Apache 2.0 licence. Mistral's stated design target for the family is edge deployment across a wide hardware range, and this is the variant that reaches the low end of it — the sizes that run on a phone-class accelerator or a Jetson board rather than a desktop GPU. What the size does cost is headroom: the vendor describes the 3B as its most efficient rather than its most capable model, and reserves claims like the 85% AIME 2025 score for the 14B reasoning variant. Hosted access costs 0.10 US dollars per million tokens, input and output alike, under the identifier ministral-3b-2512.

▸Videos

No videos yet.

▸Reviews