Ministral 3 14B

Mistral AI's launch blog "Introducing Mistral 3" (December 2, 2025) reveals the Ministral 3 family of three dense models at 14B, 8B, and 3B sizes, each offered in Base, Instruct, and Reasoning variants for nine total releases under Apache 2.0. The post frames Ministral models as offering the best performance-to-cost ratio in their category versus the headline Mistral Large 3. All Mistral 3 models were trained on NVIDIA Hopper GPUs, and Mistral partnered with NVIDIA, vLLM, and Red Hat to deliver NVFP4 checkpoints, optimized inference on Blackwell NVL72 systems, and deployment through TensorRT-LLM, vLLM, SGLang, llama.cpp, and Ollama. The blog notes a reasoning version of Mistral Large 3 is coming soon, though it does not isolate 14B-specific benchmark numbers.

NVIDIA's technical blog from December 2, 2025 confirms Ministral 3 dense models in 3B, 8B, and 14B sizes ship with Base, Instruct, and Reasoning variants trained on Hopper GPUs and now available via Hugging Face. Deployment targets span edge platforms including NVIDIA DGX Spark, GeForce RTX AI PC, and Jetson, alongside datacenter GB200 NVL72 systems. The post details framework support across TensorRT-LLM, vLLM, SGLang, llama.cpp, and Ollama for running Ministral 3 across NVIDIA hardware tiers. It separately documents NVFP4 quantization and NVIDIA Dynamo prefill/decode disaggregation optimizations applied to Mistral Large 3, which represents family-level rather than 14B-specific deployment context.

Together AI describes Ministral 3 14B Instruct 2512 as Mistral AI's frontier 14B-class multimodal assistant, built on a 13.5B language core paired with a 0.4B vision encoder for unified text and image reasoning. The model pairs a 256K token context with strong system-prompt adherence, targeting long-horizon agents, complex chat experiences, and analytical copilots under Apache 2.0. The page positions the model for retrieval-heavy, multi-step workflows and native function calling with structured outputs, and lists Together AI quickstart guides for RAG, agents, and Next.js chat integrations. Benchmark cells for the model itself are unpopulated, with only a 71.2% FrontierMath Tier 4 figure shown alongside competitor rows using unverifiable version labels.

Mistral AI's official documentation announces Ministral 3 14B (version 25-12) reaching general availability on December 2, 2025. The model ships under an Apache 2.0 license with a 256k token context window, multimodal support, and Mistral API pricing of $0.2 per million tokens for both input and output, positioned as the largest in the Ministral 3 family with performance comparable to Mistral Small 3.2 24B. The docs page explicitly names the model key "ministral-14b-2512" and lists supported features including structured outputs, function calling, document Q&A, prefix caching, chat completions, and the /v1/batch endpoint. Release is tied to Mistral's official APIs, with the page guiding developers to the playground and comparison tooling for evaluation.