Ministral 3 14B

On December 2, 2025, Mistral announced the Mistral 3 generation, which includes three dense Ministral 3 small models (14B, 8B, and 3B) alongside Mistral Large 3, all released under Apache 2.0. Mistral framed the Ministral 3 family as offering the best performance-to-cost ratio in its category for developers. Mistral partnered with NVIDIA, vLLM, and Red Hat to optimize the new models, training them on NVIDIA Hopper GPUs and shipping NVFP4 checkpoints via llm-compressor for efficient inference on Blackwell NVL72 systems and 8×A100 or 8×H100 vLLM nodes. The launch emphasizes open-sourcing compressed formats to enable distributed intelligence across the community.

Mistral released Ministral 3 14B on December 2, 2025, as the largest model in the Ministral 3 family, reaching general availability under Apache 2.0 with version v25.12. The official documentation positions it as offering state-of-the-art capabilities comparable to Mistral Small 3.2 24B, optimized for local deployment across diverse hardware setups. The model supports a 256k token context window and exposes structured outputs, function calling, document QnA, prefix handling, chat completions, and batching through the /v1/chat/completions, /v1/conversations, and /v1/batch endpoints. A linked technical report and launch blog accompany the release, giving developers primary references for further integration and evaluation.