Without an internet connection, using just one consumer-grade graphics card, you can have a humanoid AI assistant on your personal computer that "watches, listens, speaks, and even proactively reminds you". This is not a concept -- MiniCPM-o 4.5 has already turned it into something that runs directly on your computer.

MiniCPM-o 4.5 is the industry's first end-to-end full-duplex omni-modal large model, released by ModelBest together with Tsinghua University. At just 9B parameters, it delivers a complete closed loop of multimodal perception plus streaming voice interaction. Downloads on Hugging Face have already broken 250,000+.

What Is MiniCPM-o 4.5

Simply put, it is an AI assistant model that can "watch video, listen to sound, talk back, and remind you proactively" all at the same time. Its biggest differences from traditional large models:

• Full-duplex interaction: you do not have to wait for it to finish before speaking -- you can interrupt or cut in at any moment, and it senses it in real time
• Omni-modal: video, audio, and text input plus text and speech output, all handled by a single model
• Works offline: everything runs locally, and your data never leaves your computer

How to Try It

Option 1: Online demo (zero barrier)

No registration, no downloads -- just open your browser:

• Desktop: https://minicpmo45.modelbest.cn/
• Mobile (recommended): https://minicpmo45.modelbest.cn/mobile/

The demo showcases three interaction modes -- traditional turn-based chat, full-duplex voice interaction, and full-duplex video interaction -- with full control over parameters such as the Prompt and reference audio.

Option 2: One-click installer Comni (recommended)

ModelBest provides Windows/macOS desktop installers that bundle model download, environment setup, and demo runtime.

Windows version:

• GitHub: https://github.com/tc-mb/llama.cpp-omni/releases/latest/download/Comni-Setup-win64.exe
• ModelScope: https://modelscope.cn/models/OpenBMB/MiniCPM-o-4_5-gguf/resolve/master/app/Comni-Windows-x64.exe
• Hardware requirement: a GPU with 12GB+ VRAM (RTX 4080/4090/5070/5080/5090)

macOS version:

• GitHub: https://github.com/tc-mb/llama.cpp-omni/releases/latest/download/Comni-macOS-arm64.dmg
• ModelScope: https://modelscope.cn/models/OpenBMB/MiniCPM-o-4_5-gguf/resolve/master/app/Comni-macOS-arm64.dmg
• Hardware requirement: M1-M5 Max/M5 Pro, 16GB+ of memory recommended

Option 3: Free API

API endpoint: https://api.modelbest.cn/minicpmo45/v1/

It is currently free and open, supports omni-modal full-duplex real-time interaction, and needs no VAD mechanism to control dialogue turns. API documentation: https://api.modelbest.cn/minicpmo45/docs

Option 4: Deploy from the open-source code

The demo's front-end and back-end code is fully open source: https://github.com/OpenBMB/MiniCPM-o-Demo

Linux users can clone the repository and deploy the complete demo service -- one of the first locally deployable full-duplex omni-modal interaction demo projects.

Core Technology: The Omni-Flow Streaming Omni-Modal Framework

MiniCPM-o 4.5's core breakthrough is the Omni-Flow streaming omni-modal framework. It creates a shared millisecond-level timeline that precisely aligns every information stream -- vision, audio, language -- onto time slices. The model refreshes its awareness of the environment at a high frequency of once per second and autonomously decides when to speak and when to remind.

The model architecture consists of four major modules:

• Vision encoder (0.4B): SigLIP-ViT, in charge of "seeing"
• Audio encoder (0.3B): Whisper-Medium, in charge of "listening"
• LLM base (8B): Qwen3-8B, in charge of "thinking"
• Speech token decoder (~0.3B): a lightweight Llama architecture that converts text into speech

Performance

Do not be fooled by the 9B parameter count -- MiniCPM-o 4.5 delivers standout results on several fronts:

• Vision: an OpenCompass composite score of 77.6 and an MMBench English score of 87.6, on par with Gemini 2.5 Flash
• Omni-modal interaction: fully surpasses Gemini 2.5 Flash and Qwen3-Omni on benchmarks such as Daily-Omni and Video-Holmes
• Inference efficiency: the INT4 quantized version needs only 12GB of VRAM and decodes at 212 tokens/s, 40%+ faster than Qwen3
• Speech generation: Chinese and English speech quality and emotional expressiveness both beat CosyVoice2

Application Scenarios

MiniCPM-o 4.5's full-duplex capability unlocks a series of scenarios that traditional turn-based dialogue models cannot handle:

• Proactive companion: real-time guidance and reminders while you cook, repair things, or exercise
• Accessibility assistance: serving as the "eyes" of visually impaired people, continuously observing the environment and proactively announcing key information
• Smart cockpit: continuously monitoring road conditions and the driver's state, proactively raising safety warnings
• Embodied intelligence: acting as the robot's "brain", continuously sensing a dynamic environment and making its own decisions

Related Links

• Technical report: https://github.com/OpenBMB/MiniCPM-o/blob/main/docs/MiniCPM_o_45_technical_report.pdf
• Hugging Face: https://huggingface.co/openbmb/MiniCPM-o-4_5
• ModelScope: https://www.modelscope.cn/models/OpenBMB/MiniCPM-o-4_5
• Demo repository: https://github.com/OpenBMB/MiniCPM-o-Demo