MiniCPM-o 4.5

MiniCPM-o 4.5

Handles images, video, speech and live audio-video streams in one 9B model, answering in text and speech in English and Chinese. For developers building voice and vision assistants.

Made by openbmb on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiniCPM-o 4.5 is a 9B-parameter multimodal model from openbmb. It reads images, video and speech, and it can answer in text or in speech. It joins SigLip2, Whisper-medium, CosyVoice2 and Qwen3-8B into one end-to-end model.

Who it suits: Developers building assistants that see, listen and talk, and researchers who test vision and speech models. The card says it can take live video and audio while it speaks. Use voice cloning only with voices you have the right to use.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Software | Assistant builders | Build a voice and vision assistant that runs on local hardware | A prototype answers spoken questions about what the camera sees
Document work | Back-office developers | Parse English documents from page images | A script turns a scanned page into text
Accessibility | App makers | Describe a live scene aloud | An app speaks a short description of the camera view
Research | ML researchers | Study full-duplex multimodal streaming | Compare instruct and thinking modes on the same prompts

Skip it if you only need plain text chat, because a text-only model is lighter to run.

Where to use it:

• PyTorch on an Nvidia GPU, which the card recommends for full precision
• llama.cpp and Ollama, for CPU use on local devices
• vLLM and SGLang, for high-throughput serving
• The card's open web demos on a GPU or a Mac

Use case | Who | What they get back
Image and video questions | Developers | Text answers about pictures and clips
Spoken conversation | App makers | Replies in speech, in English and Chinese, with configurable voices
Live streaming | Researchers | Text and speech produced while video and audio keep arriving

How to set it up: You need Python and pip to download the files. Running it takes an Nvidia GPU for full precision, or a smaller quantized copy for local devices.

• Python and pip
• An Nvidia GPU for PyTorch use, or a runner such as llama.cpp for CPU use
• Disk space for the weights, about 18.7 GB in BF16

1. Install the Hugging Face hub library.
2. Download the repository.
3. Follow the card's usage steps for your chosen runner.

pip install -U huggingface_hub
hf download openbmb/MiniCPM-o-4_5

How to use it: Pick a route from the card. PyTorch on an Nvidia GPU is the basic route, while quantized copies in 16 sizes suit local devices. The card also open-sourced web demos for the live streaming experience.

Tool | What it does
PyTorch inference | Runs the model at full precision on an Nvidia GPU
llama.cpp and Ollama | Run quantized copies efficiently on a CPU
vLLM and SGLang | Serve the model to many requests with less memory
Web demo | Gives a live streaming experience on a local GPU or Mac

Example: You open the local web demo and ask by voice what an object on your desk is. The model can look through the camera, listen, and answer in speech while the stream continues.

For newcomers and experts: If you are new, think of it as one model that can see a picture, hear you and talk back. Start with the card's web demo or a quantized copy before the full-size weights.

For people in the field, the encoders and decoders connect to the language model through hidden states instead of separate tools. The card describes a full-duplex streaming mechanism with an interleaved speech token decoder. The card gives an OpenCompass average of 77.6, a limit of 1.8 million pixels per image, and a ceiling of 10 frames per second for video.

Video, audio and text arrive
Encoders feed the language model
The model reasons over all inputs
Text and speech stream back

Good to know:

Title: Weight size in GB, 4-bit is Hugging Face's estimate
BF16: 18.7 GB
4-bit: 4.7 GB

• License: Apache 2.0
• Cost to run: none named, though full precision needs an Nvidia GPU
• Data it reaches: it downloads files from Hugging Face, and live mode reads whatever camera or microphone input you give it
• Not for: cloning a voice you have no right to use
• The card supports more than 30 languages, with speech in English and Chinese

On Hugging Face, 13 November 2026

• Task: any to any
• Library: transformers
• Parameters: 9.4B
• Weights: about 18.7 GB in BF16; about 4.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 170,793
• Likes: 1,571
• License: apache-2.0
• Quantized versions: 13
• Fine-tunes: 9
• Tags: transformers, onnx, safetensors, minicpmo, feature-extraction, minicpm-o, minicpm-v, multimodal, full-duplex, any-to-any

Links

• Model: https://huggingface.co/openbmb/MiniCPM-o-4_5
• Files: https://huggingface.co/openbmb/MiniCPM-o-4_5/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/MiniCPM-o-4_5

Then follow the usage in its model card. Nothing in this kit runs by itself.