MiniCPM-o 4.5

OpenBMB's 9B model that sees, listens and speaks at once, with full-duplex live streaming, English and Chinese voices and Apache 2.0.

In short. MiniCPM-o 4.5 is an AI model you can talk to while it watches a camera feed. It listens and speaks at the same time instead of taking turns, and it can speak up on its own, for example to remind you of something it sees. It runs on a 12 GB GPU or a recent Mac when quantized.

Summary

MiniCPM-o 4.5 is OpenBMB's omni model that sees, listens and speaks at the same time. It is built end to end from a SigLIP2 vision encoder, Whisper-medium for audio, CosyVoice2 for speech output and a Qwen3-8B language model, with 9.37B parameters in all. It reads images of up to 1.8 million pixels, video of up to 10 frames per second and audio, and it answers in text and speech. The card describes real-time conversation in English and Chinese with configurable voices, voice cloning and role play from a short reference clip, support for more than 30 languages, and instruct and thinking modes in one model. Its headline feature is full-duplex live streaming: it takes continuous video and audio while it writes text and speech, and it decides once per second whether to speak, so it can start a reminder or comment on its own. The technical report is arXiv 2604.27393, dated 30 April 2026, and names the streaming framework Omni-Flow.

OpenBMB reports an average of 77.6 on OpenCompass, which it says is above GPT-4o and Gemini 2.0 Pro and close to Gemini 2.5 Flash. It also says the model reaches state of the art for end-to-end English document parsing on OmniDocBench, ahead of Gemini 3 Flash, GPT-5 and DeepSeek-OCR 2. These are the developer's own numbers, and the card marks which rival scores came from its own runs. The card's efficiency table lists 154.3 tokens per second and 19.0 GB of GPU memory in BF16, and 212.3 tokens per second and 11.0 GB in int4, while the same table shows Qwen3-Omni-30B-A3B using 20.3 GB in int4 and running out of memory in BF16 on its test setup.

Both the weights and the code are under Apache 2.0. The BF16 checkpoint is 18.8 GB, an AWQ int4 build is 8.5 GB, and the GGUF language-model files run from about 5 GB at Q4_K_M to 16.4 GB at F16, with the vision, audio and speech parts in separate folders. The card recommends PyTorch on an NVIDIA GPU with transformers 4.51.0, and its full-duplex web demo needs a GPU with at least 28 GB. For llama.cpp-omni it lists an Apple M3, M4 or M5 with 16 GB of memory or an NVIDIA GPU with 12 GB for half-duplex speech chat, and an M4 Max with 24 GB or a 12 GB GPU for full-duplex streaming. vLLM, SGLang and Ollama are listed too. The config sets 40,960 positions. In the 30 days to 11 November 2026, Hugging Face counted about 165,000 downloads for the main repository and 118,000 for the GGUF repository, and the repository first appeared there on 3 February 2026.

Good for

  • A local assistant that watches a camera or screen and talks back in real time, in English or Chinese
  • Voice assistants and role play with custom voices, where cloning a voice from a short clip is allowed and you have consent
  • Reading photos, documents and video on one 12 to 24 GB GPU or a recent Mac
  • Commercial products, since both weights and code are Apache 2.0

Watch out for

  • The card describes speech conversation in English and Chinese only, and its claim of more than 30 languages is not specific to speech
  • The OpenCompass, OmniDocBench and speed comparisons are OpenBMB's own, and some rivals are run on its own setup
  • The full-duplex demo needs a 28 GB NVIDIA GPU, the card pins transformers to 4.51.0, and quantized builds can lose some accuracy
  • The card describes cloning a voice from a reference clip but mentions no watermark, so consent and labeling are up to you

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the 9B AWQ (int4) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
9B (BF16) BF16 9.37B 22 GB 11 GB 19 GB Gaming GPU
9B AWQ (int4) INT4 9.37B 22 GB 11 GB 11 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 11 GB of GPU memory.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is MiniCPM-o 4.5?
MiniCPM-o 4.5 is an AI model you can talk to while it watches a camera feed. It listens and speaks at the same time instead of taking turns, and it can speak up on its own, for example to remind you of something it sees. It runs on a 12 GB GPU or a recent Mac when quantized.
Can I use MiniCPM-o 4.5 commercially?
Yes. MiniCPM-o 4.5 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is MiniCPM-o 4.5 open source?
MiniCPM-o 4.5 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run MiniCPM-o 4.5 on my own computer?
Yes. The smallest variant, 9B AWQ (int4), needs about 11 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was MiniCPM-o 4.5 released?
MiniCPM-o 4.5 was released by OpenBMB on February 3, 2026.
How long a context window does MiniCPM-o 4.5 have?
MiniCPM-o 4.5 supports a context window of up to 40K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MiniCPM-V 4.7 OpenBMB 1.55B – 35.2B 3B active 256K Apache 2.0 (Permissive) 1.07 GB Laptop or phone
Nemotron 3 Nano Omni NVIDIA 31B 3B active 256K NVIDIA Open Model (Permissive) 32 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
MiMo-V2.6 Xiaomi 9.4B – 1.02T 42B active 1M MIT (Permissive) 6.49 GB Laptop or phone