OpenBMB’s AI Cat Cam Named The Right Cat In All 89 Drinks It Logged

OpenBMB‘s MiniCPM-V AI models powered a homemade cat camera that logged 89 drinks from four cats over three days and named the right cat every time, using two small ESP32 boards and a MacBook.
Key Takeaways
OpenBMB posted an open-source cat monitor build using two ESP32 boards and a MacBook to recognize cats and log drinking events
The repo describes a tiered pipeline where a local MiniCPM-V 1B model filters frames and a cloud MiniCPM-V 35B handles identity confirmation
The camera detects a cat occupying much of the frame near the bowl and then triggers follow-up checks instead of detecting water
The system stores event times, confidence scores and MP4 clips in SQLite-backed logs and shows daily counts on a small dashboard
OpenBMB posted the project on Thursday and linked to a public repo by developer emwstudio. The build uses one ESP32 camera board for video capture and one AMOLED ESP32 board for a live dashboard.
OpenBMB develops small multimodal models, including MiniCPM-V, that can run vision tasks with less hardware than frontier cloud systems.
MiniCPM-V Cat Monitor Turns Toy Hardware Into A Vision Stack
The most interesting part is not the cat theme. It is the system design.
The repository shows a tiered pipeline, a local MiniCPM-V 1B model handles the cheap first pass and asks a simple question, is there a cat in frame, while a larger cloud model, MiniCPM-V 35B, checks harder cases and handles identity confirmation against reference images.
That split cuts unnecessary calls to the expensive model and keeps the always-on sensing job on low-cost hardware while pushing only the uncertain cases to the cloud.
The Repo Shows Where Edge AI Actually Saves Money
The project notes say the camera does not detect water directly. Instead it looks for a cat occupying a large enough share of the frame near the bowl, then triggers follow-up checks.
That choice matters because reliable home automation usually fails on overcomplicated sensing, so the build uses background difference, simple thresholds and repeated positive votes before it records an event.
The large model enters only after the system already believes something worth checking happened.
The result is less like a chatbot on a gadget and more like a layered monitoring system where small models watch all the time and big models judge rarer moments.
Also Read: Anthropic Gives Away 4 Of Its AI Agent Tools For Any Developer To Use
From Cloud-Only Demos To Layered Household AI
A year ago, many multimodal demos assumed a constant cloud round trip, which made them expensive, delayed and brittle once cameras stayed on all day. This build takes a more practical route.
It stores event times, confidence scores and MP4 clips in SQLite-backed logs, then surfaces daily counts on a small dashboard. That is the kind of plumbing that turns an AI demo into a usable device.
The broader lesson is that edge AI is getting defined by routing, not just model size, the cheaper breakthrough is not squeezing a model onto a board but deciding which frames should trigger a larger model call.
How OpenBMB’s Two-Stage Vision Stack Cuts Cost
In the emwstudio design, the first stage uses MiniCPM-V-4.7-1B on a local MacBook through llama.cpp quantization.
That model checks low-cost cues, mainly whether a cat fills enough of the frame. Only after several positive checks does the system open an event.
Then MiniCPM-V-4.7-35B-A3B in the cloud reviews event frames, compares them with each cat’s reference image and decides whether the scene shows drinking.
For OpenBMB, that architecture shows why small multimodal models matter, a home camera runs all day, so the expensive part is not intelligence alone but how often intelligence gets invoked. By letting the 1B model reject most frames, the system lowers cloud spend and bandwidth use while keeping a stronger model available for identity and final judgment.
That tradeoff will matter to device makers more than benchmark scores, and the next thing to watch is whether OpenBMB packages this routing pattern into reference software for other camera tasks like elder care, inventory checks or shop floor monitoring.
Read Next: Anthropic Gives Buyers Proof A Claude Reply Was Not Swapped