TaoMate-H3 is a low-latency streaming audio-video generation runtime built on MiniMax H3 with a 3-step LoRA. It generates synchronized audio and video in small chunks and supports continuous long-form generation at 480p/768p/1080p resolutions.

Developed by the Alibaba TaoLive AIGC Team. Powered by MiniMax H3.

• 🤩2026/09/08: We first migrate MiniMax H3 to TaoMate (T2AV) for streaming. The new models of complete streaming FL2AV and Ref2AV are on the way. Please stay tuned!❤️

• Three-step LoRA streaming generation — each small chunk uses three Stage3 denoising intervals.
• Audio-video joint generation — speech, sound, and video are generated on one synchronized timeline.
• Low chunk latency — each small chunk reaches its final latent state well before a full MiniMax H3 request completes.
• Long-form continuity — clean KV cache and integrated audio guidance preserve visual identity, voice, and motion across prompt boundaries.
• Multiple resolutions — portrait and landscape generation at 480p, 768p, and aligned 1080p.
• Single-node inference — supports 4 GPUs or 8 GPUs with TP2 and Ulysses sequence parallelism.

• Linux
• Python 3.10 or 3.11
• NVIDIA Hopper/SM90 GPUs; 8 × H20 96 GB is the validated configuration
• CUDA 12.8 and PyTorch 2.8
• FFmpeg with H.264 and AAC support

Create the environment:

conda create -n taomate-h3 python=3.10 -y
conda activate taomate-h3

pip install torch==2.8.0 torchvision==0.23.0 \
  --index-url https://download.pytorch.org/whl/cu128
pip install triton==3.4.0 vllm==0.11.1

git clone https://github.com/Dao-AILab/flash-attention.git
pip install --no-build-isolation ./flash-attention/hopper

git clone https://github.com/TaoLiveAIGC/TaoMate-H3.git
cd TaoMate-H3
pip install -e .

Install FFmpeg on Ubuntu or Debian:

sudo apt-get update
sudo apt-get install -y ffmpeg

Download the MiniMax H3 FL2VA model:

hf download MiniMaxAI/MiniMax-H3 \
  --include "model_index.json" "FL2VA/*" \
  --local-dir models/MiniMax-H3

TaoMate-H3 LoRA weights are available on Hugging Face. The current release is the step-3000 generator EMA adapter (rank 128, alpha 128). Inference downloads it automatically to models/TaoMate-H3 on first use. To download it in advance:

hf download TaoLiveAIGC/TaoMate-H3 \
  --include "config.json" "adapter_config.json" "adapter_model.safetensors" \
  --local-dir models/TaoMate-H3

For private or gated models, run hf auth login with an account that has access. Downloads also accept the standard HF_TOKEN environment variable.

The LoRA directory contains:

models/TaoMate-H3/
├── config.json
├── adapter_config.json
└── adapter_model.safetensors

TaoMate-H3 accepts either one prompt through --prompt or one prompt per five-second block through --prompt-json.

Example prompt file:

{
  "prompts": [
    "prompt for seconds 0-5",
    "prompt for seconds 5-10"
  ],
  "seeds": [8301, 8301]
}

Run TaoMate-H3:

python -m taomate_h3 \
  --model-root models/MiniMax-H3 \
  --prompt-json examples/prompts_10s.json \
  --duration 10 \
  --resolution 768x1376 \
  --gpus 8 \
  --devices 0,1,2,3,4,5,6,7 \
  --seed 8301 \
  --output outputs/demo_10s

The command runs the complete pipeline and writes the final video to outputs/demo_10s/video.mp4. It starts its own local distributed workers, so no external torchrun command is needed.

To use a different local LoRA, add --adapter /path/to/adapter.

Common resolutions:

For an exact 1080-pixel delivery edge, crop the generated 1088-pixel edge after inference.

--prompt and --prompt-json are mutually exclusive.

The following results were measured on one 8 × NVIDIA H20 96 GB node with TP2 × Ulysses4, a 480x864 canvas, a 10-second output, and seed 8301.

Pure DiT time excludes model loading, text encoding, VAE decoding, and media encoding. First playable video includes Video VAE decoding and H.264 publication in the matched first-chunk publication benchmark. The table covers the Stage3 generation path and excludes the command's internal audio preparation. A 10-second TaoMate-H3 run contains 24 generation forwards and eight clean-KV updates.

TaoMate-H3 is released under the MiniMax H3 Community License Agreement. Use and distribution must follow the terms of that license.

If you find our work helpful, feel free to cite us.

@misc{taomate2026,
      title={TaoMate: Anchor-Guided Memory Bridging Evolving and Reference States for Real-Time Audio-Video Digital Human Generation}, 
      author={Qijun Gan and Chenwei Zhang and Meiguang Jin and Junfeng Ma and Qiu Shen},
      year={2026},
      eprint={2607.24359},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.24359}, 
}

We thank the teams and contributors behind:

• MiniMax H3
• Qwen3-VL
• PyTorch
• FlashAttention
• Triton
• vLLM