MiniCPM-o 4.5 (9B full-duplex omnimodal)
openbmb/MiniCPM-o-4_5 is a compact “omnimodal” model that tries to make real-time interaction feel more like a call and less like turn-taking chat. The headline feature is full-duplex streaming: the model can process continuous audio/video input while generating text and speech output at the same time (no “wait for the user to stop talking” behavior). The model card also claims proactive interaction — the ability to comment or remind based on what it’s seeing and hearing.
What’s especially notable is the packaging and deployment story. It’s described as an end-to-end stack built on top of components like SigLip2 (vision), Whisper (ASR), CosyVoice2 (TTS), and a Qwen3-8B text backbone. And unlike many multimodal releases, it’s explicitly positioned to run locally: there are guides for llama.cpp and Ollama, plus quantized variants (including GGUF) and higher-throughput options like vLLM and SGLang. A good first experiment is to try the WebRTC streaming demo to understand latency and interruption behavior, then replicate the same “talk over it” scenario locally with a smaller quant.
Quick stats from the listing feed: pipeline: any-to-any · 42 likes · 0 downloads.
Source listing: https://huggingface.co/models?sort=modified