Is TaoMate-HD Preview state of the art?Pending benchmarks.Announced Nov 2, 2026; no independent results yet.
Alibaba TaoLive AIGC team's streaming audio-video model: speech and high-resolution video are generated together in short chunks, two coarse 480p steps then one 768p step each. Preview LoRA weights on MiniMax H3 FL2VA, with inference code.
The details so far
- Lab
- Alibaba / Qwen
- Announced
- Nov 2, 2026
- API
- Not yet · None; self-hosted
- Open weights
- Published · Preview LoRA on Hugging Face TaoLiveAIGC/TaoMate-HD-preview, not gated; code at github.com/TaoLiveAIGC/TaoMate-HD
- Parameters
- LoRA adapter, r=128 (620,232,704 F32 parameters); requires the MiniMax H3 FL2VA base model
- Type
- Progressive-resolution autoregressive audio-video generation (joint speech and video, streamed chunk by chunk)
- Team
- Alibaba TaoLive AIGC Team
- Base model
- MiniMax H3 FL2VA (MiniMaxAI/MiniMax-H3, FL2VA folder), downloaded separately
- Method
- Each chunk takes two denoising steps at 480p for coarse structure and one at 768p for detail, with audio generated throughout; each stage keeps its own audio-video history for the next chunk
- Output
- 768x1376 example at 24 FPS with 32 kHz stereo audio; prompts cover five-second windows and can be continued with new dialogue
- Hardware
- Inference code targets eight GPUs (serial or pipelined accelerated mode); the 10-second example was run on eight H20 96 GB GPUs
- License
- MiniMax H3 Community License Agreement, covering base and adapter (card license: other)
- Paper
- arXiv 2611.03394, v1 Nov 2, 2026
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: progressive resolution is 1.60x faster than direct 768p generation with matched BF16/SDPA execution, including audio-video decoding.
Lab claim: the optimized eight-GPU deployment adds a further 1.92x over serial progressive generation on H20 GPUs.
Lab claim: continuation training lowers character error rate from 30.75% to 19.57% (11.18 points), and the paper demonstrates hour-long audio-video streams.
Sources
PAPERTaoMate-HD: Progressive-Resolution Autoregressive Audio-Video Generation (arXiv 2611.03394)arxiv.org/abs/2611.03394 · Nov 2, 2026LABTaoMate-HD project pagetaoliveaigc.github.io/TaoMate-HD/ · Nov 3, 2026REPOTaoLiveAIGC/TaoMate-HD-preview on Hugging Facehuggingface.co/TaoLiveAIGC/TaoMate-HD-previewREPOTaoLiveAIGC/TaoMate-HD on GitHubgithub.com/TaoLiveAIGC/TaoMate-HDTracking continues until an independent board lists it.