News

Alibaba Releases TaoMate-H3: Streaming MiniMax H3

Compiled by Sergey KostenchukPublished 2026-09-13Updated 2026-09-18

2026-09-13 Research NVIDIA

Expanded analysis for this story

Open the longform version with context, source trail, and what changed.

TaoMate-H3 — low-latency streaming audio-video generation runtime (Alibaba TaoLive AIGC) Source

🎬 TaoMate-H3: Alibaba moves into streaming with MiniMax H3

Alibaba's TaoLive AIGC team has released TaoMate-H3: video and audio are generated in chunks at 480p/768p/1080p. On an 8×NVIDIA H20 96GB node, the first 10-second clip appears in 17.3 seconds, compared to 183.3 seconds for batch H3.

🌍 The focus shifts from the model to inference: distillation down to 4 steps (12.5x fewer transformer computations), int8/W4A8 quantization, and multi-GPU parallelism TP2 + Ulysses. The cost per second of video drops.

👤 FastH3 Live handles a continuous real-time video+audio stream on a single RTX 5090 (448×448, 22.1 fps). For 1080p, an 8×NVIDIA H20 96GB node is required. In ComfyUI — outpainting, camera control, and turbo-LoRA.

Source 1: https://github.com/TaoLiveAIGC/TaoMate-H3 Source 2: https://huggingface.co/datasets/jacokon/fasth3-live

Sources

Story

Alibaba's TaoLive AIGC team has released TaoMate-H3, a streaming runtime based on the 33B-parameter MiniMax H3 diffusion transformer, which produces a first 10-second clip in 17.3 seconds on an 8×NVIDIA H20 96GB node, compared to 183.3 seconds for the original H3. The MiniMax H3 ecosystem is shifting from the model itself to inference economics: 4-step distillation (FastVideo-Minimax-FastH3-Preview-v0.2, 12.5x fewer transformer computations compared to 50-step sampling), int8/W4A8 quantization, and multi-GPU parallelism (TP2 + Ulysses) drastically reduce the cost per second of generated video.