What is MiniCPM-o 4.5?
MiniCPM-o 4.5 is OpenBMB's open omni model, released in February 2026. It joins four existing parts end to end: a SigLIP2 vision encoder, a Whisper-medium audio encoder, a CosyVoice2 speech decoder and a Qwen3-8B language model, for about 9 billion parameters in total. It reads images up to 1.8 million pixels and video at up to 10 frames a second, hears speech, and answers in text or a spoken voice that can be cloned from a short reference clip.
The headline feature is full duplex: the model takes in continuous video and audio while it is speaking, so it can be interrupted, and it decides once a second whether to say something unprompted, such as a reminder about what it sees. Speech conversation works in English and Chinese. One checkpoint holds both an instruct and a thinking mode.
The weights are under Apache 2.0. In the last 30 days they drew about 191,000 Hugging Face downloads. It is a different line from the vision-only MiniCPM-V 4.7, which OpenBMB released in October.
Is MiniCPM-o 4.5 worth it?
MiniCPM-o 4.5 puts live voice, video and vision in one open 9B model, and its vision scores hold up against closed models. On OpenBMB's card it averages 77.6 on OpenCompass's eight vision tests, ahead of GPT-4o at 75.4 and Qwen3-VL-8B at 76.5 and just behind Gemini 2.5 Flash at 78.5. It scores 87.6 on MMBench English, 80.1 on MathVista and 94.7 on DocVQA, and OpenBMB claims the top end-to-end score for English document parsing on OmniDocBench. All of these are self-reported.
The speed table explains why it suits local use. In BF16 it decodes at 154.3 tokens a second with 0.6 seconds to first token in 19.0 GB of GPU memory; in int4 that becomes 212.3 tokens a second in 11.0 GB. Qwen3-Omni-30B-A3B, its nearest open rival, ran out of memory in BF16 on the same test and needed 20.3 GB in int4. Text skills hold up for a 9B model: 77.0 on MMLU, 86.6 on HumanEval and 94.5 on GSM8K.
The limits are in the speech side. Spoken conversation covers only English and Chinese, although text works in more than 30 languages. The full real-time web demo asks for an NVIDIA GPU with at least 28 GB, and the cheaper llama.cpp-omni route still needs an Apple M4 Max with 24 GB or a 12 GB GPU for full duplex. Voice cloning from a reference clip is built in, which is useful for products and a consent problem you have to manage yourself.
A phone-sized assistant this is not, despite the card's tagline. It is a desktop or single-GPU model, and the newer MiniCPM-V 4.7 is the better choice if you only need to read images and documents.
Where it earns its place
A robotics engineer at a ten-person home-assistant start-up
Give a prototype a voice that can see the room and be interrupted mid-sentence
Run the int4 build on a 12 GB GPU through llama.cpp-omni and stream camera and microphone input. OpenBMB measures 11.0 GB and 0.6 seconds to first token; the 1 Hz speak-or-wait decision lets it raise alerts unprompted.
A product manager at a Chinese-language tutoring app
Prototype a spoken tutor that reads a student's worksheet from the camera
Use the bilingual speech mode with a configurable voice and pass worksheet photos up to 1.8 million pixels. Its 80.1 on MathVista suits maths homework; budget a 28 GB GPU for the full web demo.
An operations analyst at a logistics firm
Pull fields from scanned English delivery notes without a cloud API
Serve the BF16 model with vLLM on a 24 GB card and prompt for structured output. The 94.7 DocVQA score and the OmniDocBench claim are self-reported, so check a sample of extractions before relying on them.
Before you commit
- Setup
- Load with transformers on an NVIDIA GPU for full precision, or use the GGUF and int4 builds in llama.cpp, llama.cpp-omni or Ollama. OpenBMB ships a Docker image for the full-duplex web demo on a Mac.
- Integrations
- vLLM and SGLang for serving, llama.cpp and Ollama for CPU and Mac, FlagOS for other AI chips. The web demo can act as an API backend for other apps.
- Data and privacy
- Self-hosted, so camera, microphone and documents stay on your hardware. Apache 2.0 permits commercial use and modification. Voice cloning from reference audio is a built-in feature.
- Cost at scale
- Weights are free. Plan on 19.0 GB of GPU memory in BF16 or 11.0 GB in int4, per OpenBMB's table; no hosted per-token price is published.
- Limits
- Speech conversation in English and Chinese only; half-duplex speech in the main web demo is still under development. The context window is not published on the card.
What’s good
- Full-duplex video, audio and speech in one 9B model
- 77.6 OpenCompass average, above GPT-4o (self-reported)
- 11.0 GB and 212.3 tokens a second in int4
- Apache 2.0 weights with llama.cpp and Ollama builds
What’s not
- Spoken conversation in English and Chinese only
- Full real-time demo needs a 28 GB GPU
- Voice cloning raises consent questions you must handle
- Vision-only MiniCPM-V 4.7 is newer for document work
Best for
Developers building a private, local voice-and-camera assistant in English or Chinese on one GPU or a high-end Mac.
MiniCPM-o 4.5 alternatives
MiniCPM-o 4.5: common questions
- Is MiniCPM-o 4.5 worth it?
- This is the open model to start from if you want a local assistant that sees and talks in real time, as long as English or Chinese is enough. We rate it 4.1 out of 5.
- How much does MiniCPM-o 4.5 cost?
- Free Apache 2.0 weights. No first-party hosted price published; you pay for your own hardware.
- What are the downsides of MiniCPM-o 4.5?
- Spoken conversation in English and Chinese only; full real-time demo needs a 28 GB GPU; voice cloning raises consent questions you must handle; vision-only MiniCPM-V 4.7 is newer for document work.
- Who is MiniCPM-o 4.5 best for?
- Developers building a private, local voice-and-camera assistant in English or Chinese on one GPU or a high-end Mac.
- How much does MiniCPM-o 4.5 cost for a team or in production?
- Weights are free. Plan on 19.0 GB of GPU memory in BF16 or 11.0 GB in int4, per OpenBMB's table; no hosted per-token price is published.
- Is MiniCPM-o 4.5 safe for company data?
- Self-hosted, so camera, microphone and documents stay on your hardware. Apache 2.0 permits commercial use and modification. Voice cloning from reference audio is a built-in feature.
- What are the best MiniCPM-o 4.5 alternatives?
- The closest alternatives we have reviewed are MiniCPM-V 4.7 (35B-A3B and 1B), Gemma 4, Phi-4-multimodal and LFM3.
Answers checked Nov 7, 2026.
Make MiniCPM-o 4.5? Show this rating on your site
