Ministral 3 14B Instruct 2512
Ministral 3 14B Instruct is Mistral’s largest “Ministral 3” checkpoint: an instruction-tuned 14B model in FP8 with a small vision encoder attached. The positioning is “edge-capable but not tiny”: it’s meant to be deployable on a single GPU while still handling serious assistant workloads (tool use, structured outputs, and long-context tasks).
Two practical details make this one worth a look. First, the context window is huge (256k), so you can use it for document-heavy chat and agentic workflows without constantly rebuilding retrieval glue. Second, it’s designed for native function calling and JSON-style outputs, which is usually where smaller/open models fall down in real applications.
What to try first: serve it with vLLM (the upstream docs recommend vLLM >= 0.12.0 with tokenizer_mode mistral / tool-call-parser mistral), then run a couple of “real” tool calls end-to-end (not just a chat demo). If you don’t need the full 256k window, cap --max-model-len early to keep memory usage sane.
Source listing: https://huggingface.co/models?sort=modified