Alibaba open-sources TaoMate-H3: 3-step streaming audio-video generation, 27x faster first segment
机器之心 · wechat · 2026-10-02
Alibaba's TaoLive AIGC team open-sourced TaoMate-H3, an audio-video joint streaming generation model built on MiniMax H3, combining 3-step LoRA inference with autoregressive generation for minute-long videos with dialogue, singing, and ambient sound. On 8×H20, DiT compute dropped from 169.6s to 14.8s (11.45x) and first-segment latency from 170s to 6.1s (27.66x). Key techniques include SelfForcing training, rolling audio-video KV cache, continuous RoPE time encoding, and audio-trajectory guidance. Code and LoRA weights are on GitHub and Hugging Face.
More from Infra
- MiniMax H3 with 360 orbit LoRA runs on 8GB VRAM: 736x576 in 7 minutes — big-boss_97 · 2026-10-02
- NVIDIA applied DL VP: at the scaling limit, efficiency is the new intelligence — ctnzr · 2026-10-02
- IBM's Torch Spyre team: an agentic CI pipeline to keep pace with PyTorch — lmoroney · 2026-10-02
- ModelHarbor: An Open-Source Self-Hosted Model Library — toks-love · 2026-10-02
- US egocentric cleaning data (100s of hrs/week capacity) can't even sell at break-even pricing — paigeinsf · 2026-10-02
- Buyers now fill out export control declarations when purchasing RTX 5090s in stores — blelbach · 2026-10-02