Mistral releases Ministral 3, edge models in 3B, 8B and 14B under Apache 2.0
Ministral 3 is a family of 3B, 8B and 14B edge models with vision, in base, instruct and reasoning versions, under Apache 2.0.
Mistral AI published Ministral 3, a family of edge models. The repository names show 3B, 8B and 14B sizes, each with base, instruct and reasoning variants. This item uses the card for Ministral 3 14B Instruct 2512, which the card calls the largest of the family and says is comparable in performance to the larger Mistral Small 3.2 24B.
The model is a language model with vision capabilities, released as an FP8 instruct version for chat and instruction use. The card says it can be deployed locally and fit in 24 GB of VRAM in FP8, and less if quantised further. Listed languages include English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic.
The licence is Apache 2.0.
The card states the sibling sizes by repository name only. The date is the Hugging Face repo creation date of the 14B instruct repo.
Practical sizes under a licence that asks nothing. The 14B-versus-24B claim is Mistral's.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Apache 2.0 weights in several sizes; the 14B FP8 fits 24 GB of VRAM, per the card.
- Useful 3.5: Edge and local chat, with vision input.
- Notable 3.5: 345 likes on the 14B instruct card.
- Novel 2.5: A size ladder with reasoning variants; not a new design.
Comments