MiniCPM-o 4.5 uses an end-to-end omni-modal architecture built from SigLip2, Whisper-medium, CosyVoice2, and Qwen3-8B, totaling 9 billion parameters, and supports full-duplex multimodal streaming interaction. Per-capability evaluation figures are given in the table below.
Highlights
77.6 Comprehensive evaluation average score
OpenCompass — Vision benchmark results：Per-benchmark comparisons against GPT-4o, Gemini 2.0 Pro, and Gemini 2.5 Flash are given in the evaluation table.
Powerful voice ability：Supports bilingual real-time voice dialogue in Chinese and English, configurable timbre, and supports voice cloning and role-playing.
Full-duplex interaction：Enables real-time video and audio input and text and voice output, and the model can actively initiate interactions.
Technical highlights
End-to-end omnimodal architecture：The vision, speech encoder/decoder and LLM are densely connected through hidden states to achieve better information flow and control and fully mine multimodal knowledge.
Full-duplex real-time streaming mechanism：Transform offline encoder/decoder into online full-duplex mode to support streaming input and output.
The speech token decoder generates text and speech tokens in an interleaved manner, enabling full-duplex speech generation (e.g., stable long speech over 1 minute).
Multimodal streams are synchronized in the LLM backbone through a time division multiplexing (TDM) mechanism, dividing parallel streams into sequential information groups with millisecond time slices.
Active interaction mechanism：LLM continuously monitors the input video and audio streams at a frequency of 1Hz, and independently decides whether to speak. The high decision-making frequency combined with the full-duplex feature enables active interaction.
Benchmarks
Visual reasoning and perception：Scores for MiniCPM-o 4.5-Thinking on visual reasoning and perception benchmarks are given in the table below.
Voice interaction capability：The voice interaction capability covers Chinese and English speech generation and long text speech generation. MiniCPM-o 4.5-Instruct maintains low levels on multiple error rate indicators.
Full-duplex real-time streaming interaction：Full-duplex real-time streaming interaction evaluation covers daily real-time interaction, world perception and video reasoning capabilities.
Document parsing：The document parsing evaluation is based on the overall edit distance of the Chinese and English subsets of OmniDocBench. The lower the value, the better the parsing quality.
speech production ability：The speech generation ability evaluation covers Chinese ASR, English ASR and English-Chinese speech translation, reflecting the comprehensive performance of speech recognition and generation.