MiniCPM-o 4.5

MiniCPM-o 4.5 is OpenBMB's omni-modal model, published on Hugging Face on 3 February 2026, and what separates it from the vision-language models around it is that it does not take turns. The model processes continuous video and audio input while generating text and speech output at the same time, without one blocking the other — what the company calls full-duplex live streaming. It can also act unprompted, raising a reminder or a comment from what it is currently watching, rather than waiting to be asked. It is a stitched system rather than a single tower: SigLip2 for vision, Whisper-medium for hearing, CosyVoice2 for speech and Qwen3-8B for language, connected end to end through hidden states rather than through text hand-offs. The total comes to 9,371,787,666 parameters. The language core runs 36 layers at hidden size 4096 with grouped-query attention (32 query heads to 8 key-value heads) over a 151,748-token vocabulary, and the configuration file declares a 40,960-token position budget. The vision tower is 27 layers at width 1152 on 980-pixel inputs; the audio encoder is 24 layers at width 1024. OpenBMB's reported results place it above its weight class. It gives an average of 77.6 on OpenCompass across eight benchmarks, and states that at 9 billion parameters this puts it past GPT-4o and Gemini 2.0 Pro and close to Gemini 2.5 Flash on vision-language work. For document parsing it claims state of the art on OmniDocBench for end-to-end English, ahead of Gemini-3 Flash, GPT-5 and the specialised DeepSeek-OCR 2. On MMHal-Bench, which measures hallucination, it is reported level with Gemini 2.5 Flash. All of these are the vendor's own numbers. The practical envelope: images up to 1.8 million pixels at any aspect ratio, video up to 10 frames per second, more than 30 languages, and bilingual real-time speech conversation in English and Chinese with configurable voices. A reference clip is enough for voice cloning, which OpenBMB reports as better than CosyVoice2 — the very component it is built on. Instruct and thinking modes live in the same weights. Distribution follows the rest of the family. The licence is Apache 2.0 with no threshold; quantised builds ship as int4 and GGUF in sixteen sizes; llama.cpp and Ollama cover local CPU inference, vLLM and SGLang cover throughput, and OpenBMB open-sourced a real-time web demo that runs the full-duplex experience on a MacBook or a single GPU. The repository records 992,822 downloads.

▸Videos

No videos yet.

▸Reviews