MiniCPM-o 4.5 实现全双工实时语音视频交互，9B 开源模型
做多模态 AI 交互的开发者终于有了可部署的全双工开源方案——MiniCPM-o 4.5 把实时语音视频对话从概念变成 9B 模型，值得直接上手试。
OpenBMB 发布 MiniCPM-o 4.5，一个 9B 参数的全双工多模态模型，能同时看、听、说。
它基于 Omni-Flow 框架，将交互视为连续时间流，打破传统轮询式对话，实现实时感知与响应。该模型在语音生成质量上超越 Qwen3-Omni-30B-A3B，且支持 12GB RAM 边缘部署。
这标志着 AI 交互层从“对讲机”模式迈向自然对话的关键一步。
Just a few days back, Thinking Machines Lab (TML), showcased a way of making AI interaction continuous instead of turn-based, a Full-Duplex Time-aligned micro-turn.
It's a preview of the future of a near-realtime AI voice and video conversation with new 'interaction models'
And MiniCPM-o 4.5 already shipped the same core idea through OpenBMB’s Omni-Flow framework: time-aligned perception and response instead of old turn-based chat.
A 9B Full-Duplex omnimodal model that can see, hear, and speak at the same time.
Omni-Flow also treats interaction as a continuous stream on a shared temporal axis, aligning visual input, audio input, and output speech/text into time chunks so the model can perceive while responding.
That breaks the old walkie-talkie UX of AI: user talks, model waits, model replies.
And this is not just a demo concept. It is a 9B open model with code, weights, a report, and edge deployment under 12GB RAM.
It also surpasses Qwen3-Omni-30B-A3B in omni-modal capabilities and speech generation quality.
This feels like the interaction layer AI was missing.
OpenBMB already shipped this as a real Full-Duplex omni-modal architecture, with video tokens, audio tokens, LLM hidden states, speech tokens, and waveform generation all synced to one shared timeline.
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include MiniCPM-o 4.5, 全双工, 多模态, 开源/仓库.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
