GitHub radar
MiniCPM-V 4.6: image and video understanding in 1.6 GB
OpenBMB's ~1B multimodal model understands photos and video clips locally — 1.6 GB download, runs on a phone or any laptop, over 50% lighter on visual encoding FLOPs than its predecessor. Quality lands at the 2B class, not at 4.5 level.
MiniCPM-V 4.6 is a pocket-sized multimodal model from OpenBMB: roughly 1 billion parameters, 1.6 GB at full precision. It is the lightest model in the MiniCPM-V line — version 4.5 carried 8 billion parameters. It reads single images, multiple images at once, and video clips, with a 256,000-token context window. The size drop comes from a different backbone: SigLIP2-400M plus Qwen3.5-0.8B in place of Qwen3-8B. Separately, mixed visual token compression cuts encoding FLOPs by more than 50%. There is a trade-off, and OpenBMB state it plainly: the model reaches Qwen3.5 2B-level capability on many benchmarks, while the larger 4.5 stays more accurate. Officially supports iOS, Android, and HarmonyOS, and runs via Ollama with a single command.
Why a vibe-coder should care
The 1.6 GB size means anyone can run it — on a phone, a basic laptop, or an older machine. You can ask an agent to look at a screenshot, document photo, or video frame without sending data to cloud services. Useful when working with private materials you don't want leaving your device.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install MiniCPM-V 4.6 by OpenBMB: ollama run minicpm-v4.6 — 1.6 GB download, runs on any laptop or phone. After setup, show me how to send it an image and ask what's in it.
A laptop or smartphone with 2 GB of free memory — the model takes 1.6 GB.
Open on Hugging Face▌ More finds