Ministral 3 14B

How it is built

The name says 14B. The card splits the model into a 13.5B language model and a 0.4B vision encoder, 13.9B in all.¹¹1Mistral AI, Ministral 3 14B Instruct 2512 model card, Hugging Face. https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512. Arithmetic: 13.5B + 0.4B = 13.9B.

It is a dense model of 40 layers with a hidden size of 5,120.²²2Mistral AI, params.json in the official repository: n_layers 40, dim 5120, n_heads 32, n_kv_heads 8, hidden_dim 16384, max_position_embeddings 262144. https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512/blob/main/params.json Each layer has 32 query heads and 8 key-value heads, and the feed-forward block is 16,384 wide.²²

Mistral did not train it from scratch. The Ministral 3 paper calls the method Cascade Distillation, "an iterative pruning and continued training with distillation technique".³³3Mistral AI, "Ministral 3", arXiv 2601.08584, January 2026. https://arxiv.org/abs/2601.08584 The 14B started as a pruned copy of Mistral Small 3.1, a 24B model that was also its teacher in pre-training.³³ In pre-training, distilling from Small 3.1 worked better than distilling from the much stronger Mistral Medium 3. Medium 3 became the teacher only in post-training.³³

The family was trained on between 1 and 3 trillion tokens.³³ Its context was stretched from 16,384 to 262,144 positions with YaRN and position-based softmax temperature scaling.³³

What is new

Ministral 3 came out on 2 December 2025. The 14B is the largest of its three sizes, and the others are 8B and 3B.⁴⁴4Mistral AI, "Introducing Mistral 3", 2 December 2025. https://mistral.ai/news/mistral-3

Every size comes as a base model and an instruction-tuned model, with a reasoning version on top. All of them read images.⁴⁴

Mistral's documentation calls its performance comparable to the larger Mistral Small 3.2 24B.⁵⁵5Mistral AI documentation, Ministral 3 14B model card, version 25.12, released 2 December 2025, prices per million tokens, read 29 September 2026. https://docs.mistral.ai/models/model-cards/ministral-3-14b-25-12 The reasoning version scores 0.850 on AIME 2025, where Qwen3-14B with thinking scores 0.737.¹¹ Its GPQA Diamond score is 0.712, and LiveCodeBench gives 0.646.¹¹

The instruct version is set against Qwen3 14B without thinking. It leads on Arena Hard, 0.551 to 0.427, and on WildBench, 68.5 to 65.1.¹¹

Running it

The instruct weights ship in FP8. The card says the model fits "in 24GB of VRAM in FP8, and less if further quantized".¹¹ The weights alone come to about 14 GB (13.9B x 1 byte), before the KV cache.⁶⁶6Arithmetic: 13.9B parameters x 1 byte in FP8 = 13.9 GB, weights only. https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 Base and reasoning versions are in BF16.¹¹

Mistral recommends a temperature below 0.1 in production. Images should have an aspect ratio close to 1:1.¹¹

Input and output cost the same on the API, $0.2 per million tokens, under the id ministral-14b-2512.⁵⁵ The context window is 256k tokens.²²⁵⁵

Every version is under Apache 2.0.⁴⁴ So is Mistral Small 4, a mixture of experts of 119B.⁷⁷7Mistral AI documentation, Mistral Small 4 model card. https://docs.mistral.ai/models/model-cards/mistral-small-4-0-26-03