MiniCPM-o 4.5 brings full-duplex live streaming to 9B open-source AI
OpenBMB just rolled out MiniCPM-o 4.5, a 9B omnimodal model that sees, listens, and speaks simultaneously in real time. Thanks to an end-to-end full-duplex design, it doesn't just wait its turn—it monitors inputs at 1Hz and can proactively interrupt with reminders. It matches Gemini 2.5 Flash on key multimodal tasks while running locally on a Mac or single GPU.
MiniCPM-o 4.5 is a new 9B open-source omnimodal model from OpenBMB that brings real-time, live-streaming voice and vision interaction directly to your local hardware.
Unlike traditional chatbots that force you to wait through rigid back-and-forth turns, MiniCPM-o 4.5 is built for true full-duplex communication. You can point your camera at something and talk naturally; the model looks, listens, and speaks all at the same time without missing a beat.
Even wilder is its proactive interrupting capability. You can interrupt the model mid-sentence, but it can also make a decision once every second to interrupt you with a reminder or observation if it spots something relevant in your video stream.
Under the hood and benchmark results
OpenBMB built MiniCPM-o 4.5 end-to-end by fusing SigLip2, Whisper-medium, CosyVoice2, and Qwen3-8B into a single 9B parameter system. The encoders, decoders, and LLM backbone share hidden representations directly instead of passing intermediate text pipelines.
To keep everything synchronized in real time, the model uses a millisecond-level time-division multiplexing mechanism. It slices parallel audio and video feeds into synchronized timeline chunks, while interleaving text and speech token decoding so speech generation stays rock-solid even beyond one minute.
It also handles reference-based voice cloning and role-play straight out of the box using a simple audio prompt, outperforming standalone tools like CosyVoice2 on naturalness.
On the OpenCompass multimodal benchmark, MiniCPM-o 4.5 scored 77.6, edging past GPT-4o and Gemini 2.0 Pro while closing in on Gemini 2.5 Flash. On OmniDocBench for end-to-end English document parsing, it outperformed proprietary giants like Gemini-3 Flash and GPT-5.
Specs, hardware, and where to run it
The full unquantized model weighs roughly 23.4GB, easily fitting onto a single Nvidia RTX 4090 GPU with PyTorch.
If you are on an everyday machine, you do not need enterprise gear. OpenBMB released 16 different quantized int4 and GGUF variants compatible with llama.cpp and Ollama, meaning you can run it locally on Apple Silicon Macs and regular CPUs. For high-throughput production, it natively supports vLLM and SGLang.
You can grab the weights and code on Hugging Face and GitHub right now, where OpenBMB has also published a deployable real-time web demo for local GPUs and Macs.
Why it matters
MiniCPM-o 4.5 breaks the monopoly proprietary frontier labs held over fluid, real-time voice and vision experiences. For developers building hardware companions, wearable camera assistants, or in-car AI, having an open 9B model that runs entirely offline with sub-second duplex feedback is a massive unlock.
We think the 1Hz proactive interruption mechanism is a stroke of engineering genius that makes conversational AI actually feel human. While sustaining continuous multi-stream inference on consumer chips still pushes memory bandwidth to its limits, this release sets a serious new bar for local edge intelligence.
Sources (1)
- huggingface.co(opens in a new tab)Originalhuggingface.co