Ministral 3 8B

Ministral 3 8B is the 8B-class member of Mistral's Ministral 3 family released in December 2025, designed as a dense transformer that can run on a single GPU and even fit in roughly 12GB of VRAM in FP8, making it well suited to local assistants, embedded chat interfaces, and other edge deployments. Unlike Mistral Small 4's mixture-of-experts approach, every Ministral 3 checkpoint is fully dense, which keeps serving infrastructure simple and reduces communication overhead for compact inference pipelines.

Architecturally, the model pairs an 8.4B language model with a 0.4B vision encoder, so it can analyze images and text together, support dozens of languages, and maintain strong adherence to system prompts for agentic workflows that rely on native function calling and JSON output. The open Apache 2.0 release and 256k context window make it a practical daily-driver for chat, translation, image and document description, and lightweight reasoning workloads where fitting a full-strength multimodal model on consumer or single-server hardware matters more than chasing the largest parameter counts.