What is your device?
│
├─ NVIDIA GPU
│   ├─ VRAM >= 28GB (A100 / H100 / RTX A6000)
│   │   └─→ Web Demo (full bf16) ✅
│   │
│   ├─ VRAM 12-24GB (RTX 3060 12G / RTX 4070 / RTX 4090)
│   │   └─→ Web Demo (AWQ/BNB int4 quantized) ✅
│   │
│   └─ VRAM < 12GB
│       └─→ ⚠ Not yet supported (llamacpp-omni / vLLM planned)
│
├─ Apple Silicon (M1/M2/M3/M4/M5 Mac)
│   └─→ ⚠ Not yet supported (llamacpp-omni planned, stay tuned)
│
├─ CPU only (no GPU)
│   └─→ ⚠ Not yet supported (llamacpp-omni planned, stay tuned)
│
└─ Other platforms (AMD GPU / Intel Arc / Ascend NPU, etc.)
    └─→ ⚠ Not tested and not yet supported

Agent guideline: Only NVIDIA GPUs with VRAM >= 12GB are currently supported (Web Demo). For all other devices — Apple Silicon, CPU-only, NVIDIA < 12GB, AMD, Intel Arc, Ascend NPU, etc. — inform the user: "This device is not yet supported for MiniCPM-o 4.5 deployment. Support for Apple Silicon and low-resource devices via llamacpp-omni is planned. Please stay tuned for updates." Do NOT attempt to guide the user through any deployment steps for unsupported devices.

Deployment Methods Overview

Method	Use Case	Model Format	Section
Web Demo	Interactive experience (chat / voice / full-duplex)	bf16 / AWQ / BNB	Web Demo Deployment
vLLM Serve	High-throughput API serving	bf16 / AWQ	vLLM Serve Deployment
llamacpp-omni	Low-resource / Apple / CPU inference	GGUF	llamacpp-omni Deployment

References

Documentation: https://minicpm-o.readthedocs.io/
HuggingFace: https://huggingface.co/openbmb/MiniCPM-o-4_5
ModelScope: https://modelscope.cn/models/OpenBMB/MiniCPM-o-4_5

Hardware Requirements

Variant	Precision	Model Size	Inference VRAM	Recommended Device
Full (bfloat16)	bf16	~18GB	~21.5GB	NVIDIA >= 28GB (A100 / H100 / RTX A6000)
AWQ quantized (int4)	W4A16	~6GB	~11GB	NVIDIA >= 12GB (RTX 3060 12G / RTX 4070)
BNB quantized (int4)	NF4	~6GB	~11GB	NVIDIA >= 12GB (RTX 3060 12G / RTX 4070)
GGUF (llama.cpp)	Q4_K_M	~6GB	~12GB VRAM or 16GB RAM	NVIDIA >= 12GB / Apple M3+ >= 16GB / CPU only

Pre-quantized AWQ model: openbmb/MiniCPM-o-4_5-awq
llama.cpp-omni full-duplex requires Apple M4 Max >= 24GB RAM or NVIDIA >= 12GB

Web Demo Deployment

Step 1: Environment Setup

1.1 Python 3.10+

Skip if Python 3.10+ is already available. Otherwise, install Miniconda:

mkdir -p ./miniconda3_install_tmp
wget https://repo.anaconda.com/miniconda/Miniconda3-py310_25.11.1-1-Linux-x86_64.sh \
    -O ./miniconda3_install_tmp/miniconda.sh
bash ./miniconda3_install_tmp/miniconda.sh -b -u -p ./miniconda3
source ./miniconda3/bin/activate

install.sh defaults to python3.10. Use PYTHON=python3.11 bash install.sh to specify another version.

1.2 FFmpeg

sudo apt update && sudo apt install -y ffmpeg

1.3 Clone Repository and Install Dependencies

git clone https://github.com/OpenBMB/MiniCPM-o-Demo.git
cd MiniCPM-o-Demo
bash ./install.sh

install.sh automatically: creates .venv/base virtual environment -> installs PyTorch 2.8.0 -> installs requirements.txt dependencies -> verifies the environment.

Manual installation alternative:

python -m venv .venv/base && source .venv/base/bin/activate
pip install "torch==2.8.0" "torchaudio==2.8.0"
pip install -r requirements.txt

Step 2: Model Download

Model size is ~18GB. Use the auto-source script to benchmark and pick the fastest source via SDK:

# Auto-benchmark sources (downloads config.json from HuggingFace / ModelScope via SDK)
python scripts/download_model.py --local-dir ./model/MiniCPM-o-4_5

# Manually specify source: huggingface / modelscope
python scripts/download_model.py --source modelscope --local-dir ./model/MiniCPM-o-4_5

The script automatically verifies the model after download. If the user already has a model, verify it separately:

# Verify a local model directory
python scripts/download_model.py --verify /path/to/MiniCPM-o-4_5

# Verify a HuggingFace Hub ID (downloads config.json to check)
python scripts/download_model.py --verify openbmb/MiniCPM-o-4_5

Checks: model_type == "minicpmo", architectures contains "MiniCPMO", version == "4.5".

Script is at scripts/download_model.py. Can also be imported: from download_model import verify_model

Manual download alternatives:

# HuggingFace CLI
huggingface-cli download openbmb/MiniCPM-o-4_5 --local-dir /path/to/MiniCPM-o-4_5

# ModelScope
modelscope download --model OpenBMB/MiniCPM-o-4_5 --local_dir /path/to/MiniCPM-o-4_5

You can also skip manual download — the model will be automatically downloaded from HuggingFace Hub on first launch (requires stable network).

Agent guideline: When the user provides a custom model path, run python scripts/download_model.py --verify <model_path> to confirm the model is valid before proceeding to configuration.

Step 3: Configuration

cp config.example.json config.json

Minimal configuration (no changes needed when using auto-download):

{
  "model": { "model_path": "openbmb/MiniCPM-o-4_5" }
}

Local model configuration:

{
  "model": { "model_path": "/path/to/MiniCPM-o-4_5" }
}

Configuration priority: CLI arguments > config.json > defaults. See web-demo-reference.md for all configuration fields.

Step 4: Generate SSL Certificate

Browser microphone/camera APIs require HTTPS. SSL certificate is mandatory:

mkdir -p certs
openssl req -x509 -newkey rsa:2048 \
    -keyout certs/key.pem -out certs/cert.pem \
    -days 365 -nodes -subj '/CN=dev'

Self-signed certificates trigger a browser security warning — click "Proceed" to continue. Replace files under certs/ when you have a proper certificate.

Step 5: Start the Service

# Single GPU
CUDA_VISIBLE_DEVICES=0 bash start_all.sh

# Multi-GPU (one Worker per GPU, parallel request processing)
CUDA_VISIBLE_DEVICES=0,1,2,3 bash start_all.sh

# HTTP mode (microphone/camera unavailable, not recommended)
bash start_all.sh --http

The startup script automatically: detects GPUs -> launches Workers (one per GPU) -> waits for model loading (~30-90s) -> starts Gateway -> prints access URLs.

Step 6: Verify and Use

Successful startup output:

Service is running!
Chat Demo:  https://localhost:8006
Admin:      https://localhost:8006/admin
API Docs:   https://localhost:8006/docs

Verify with curl:

curl -k https://localhost:8006/health

Four interaction modes:

Mode	URL	Description
Turn-based Chat	`/`	Text/image/audio/video input, streaming text + voice output
Half-Duplex Audio	`/half_duplex`	Server-side VAD, auto-detects speech start/end
Omnimodal Full-Duplex	`/omni`	Simultaneous audio + video input, model decides when to respond
Audio Full-Duplex	`/audio_duplex`	Real-time bidirectional voice conversation
Admin	`/admin`	Worker status / queue / session management

Stop the Service

kill $(cat tmp/*.pid 2>/dev/null) 2>/dev/null
# or
pkill -f "gateway.py|worker.py"

Advanced: torch.compile Acceleration

Set "service": { "compile": true } in config.json.

Pre-compile (recommended to avoid ~15min cold compilation on first run):

CUDA_VISIBLE_DEVICES=0 TORCHINDUCTOR_CACHE_DIR=./torch_compile_cache \
    .venv/base/bin/python precompile.py

With cache, startup overhead drops to ~5min. Omni Full-Duplex per-unit latency decreases from ~0.9s to ~0.5s (A100).

Troubleshooting

See web-demo-reference.md for detailed troubleshooting. Quick reference:

Issue	Resolution
OOM	Confirm VRAM >= 28GB, check usage with `nvidia-smi`
Microphone/camera unavailable	Confirm HTTPS mode and certificates exist
Slow model download	Use `--source modelscope` or ModelScope CLI
Worker fails to start	Check `tmp/worker_*.log`, verify CUDA driver / model path / port conflicts

vLLM Serve Deployment

⚠ Not Yet Available — The vLLM Serve deployment is under development and not ready for use.

Agent guideline: When the user selects or is directed to the vLLM Serve method, respond directly: "The vLLM Serve deployment is not yet available. Please use the Web Demo method as an alternative for now. We will support vLLM Serve soon — stay tuned for updates." Do NOT attempt to guide the user through vLLM deployment steps.

Planned Content Outline (In Development)

Environment setup and vLLM installation
Model loading configuration
Starting the vLLM server
API call examples (OpenAI-compatible format)
Multimodal input (image / audio / video)
Performance tuning (tensor parallel / quantization / prefix caching)
Troubleshooting

llamacpp-omni Deployment

⚠ Not Yet Available — The llamacpp-omni deployment is under development and not ready for use.

Agent guideline: When the user selects or is directed to llamacpp-omni (including Apple devices, low-VRAM NVIDIA, CPU-only, or other platforms), respond directly: "The llamacpp-omni deployment is not yet available. If you have an NVIDIA GPU (VRAM >= 12GB), you can use the Web Demo method (int4 quantized) for now. We will support llamacpp-omni soon — stay tuned for updates." Do NOT attempt to guide the user through llamacpp-omni deployment steps.

Planned Content Outline (In Development)

Environment setup and llama.cpp compilation
GGUF model conversion and download
Starting the inference service
API call examples
Multimodal input support
Quantization precision and performance comparison
Troubleshooting

minicpmo-45-deploy

MiniCPM-o 4.5 Deployment Guide

Step 0: Check Your Device and Choose a Deployment Method