MiniCPM-V & o Cookbook

Cook up amazing AI applications effortlessly with MiniCPM-V, MiniCPM-o, and the MiniCPM LLM series — bringing text, vision, speech, and live-streaming capabilities right to your fingertips.

What's new

🎉 MiniCPM-V 4.6 released — Instruct + Thinking variants, Qwen3.5 hybrid backbone, 256K context, simplified vision merger.
Inference: Single-image QA · Multi-image QA · Video · OCR · PDF · Grounding
Deployment: vLLM · SGLang · llama.cpp · Ollama
Quantization: GGUF · BNB · AWQ

Pick the right recipe

Individuals

Effortless inference on your own machine — runs on CPU + GPU, macOS / Linux / Windows, even on phones.

Ollama — easiest setup
llama.cpp — fastest CPU inference
iOS demo — runs on iPhone / iPad

Enterprises

High-throughput, scalable serving:

vLLM — production-grade GPU inference
SGLang — high-throughput GPU inference

Researchers

Train / fine-tune / customize:

Versions

This cookbook tracks all currently supported MiniCPM releases:

MiniCPM-V

Version	Status	Modality	Size	Highlights	Context
4.6 (latest)	Recommended	Image, Video	~1.2B	Phone-ready edge MLLM, LLaVA-UHD v4 vision tower	256K
4.5	Stable	Image, Video	9B	Image + video understanding, optional thinking mode	32K

MiniCPM-o

Version	Status	Modality	Size	Highlights	Context
4.5 (latest)	Recommended	Image, Video, Audio	9B	End-to-end omnimodal (vision + speech + TTS), full-duplex streaming	32K

MiniCPM (LLM)

Version	Status	Modality	Size	Highlights	Context
4.1 (latest)	Recommended	Text	8B	Hybrid reasoning, EAGLE3, InfLLM-V2	128K
4	Stable	Text	0.5B / 8B	InfLLM-V2, FRSpec speculative decoding	128K
SALA	Research	Text	8B	Sparse + linear hybrid attention	1M+

Use the version switcher in the sidebar to jump between releases.