Alibaba open-sources TLive-Omni: omni-modal model for e-commerce live streaming with 256K context
aigclink · x · 2026-08-26
Alibaba Taobao has open-sourced TLive-Omni, an omni-modal understanding model for e-commerce live streaming. Built on a Qwen3.5 backbone with an AuT audio encoder, it supports 256K token context. The model uses a Timestamped Per-vGrid layout to align audio and video tokens by time grid, solving second-level audio-visual alignment. It supports speech recognition, speaker identification, product visual grounding, OCR, temporal localization, dense video captioning, omni-modal QA, and multi-dimensional shot annotation. Versions 4B and 9B are available, and the technical report is on arXiv. Applications include building live replay systems, review tools, and content clipping.
Related event: TLive-Omni: Open-Source Omni-Modal Model for Livestream E-commerce(4 posts)→
More from Models
- China's StepFun releases Step 5 Preview weights: 600B MoE, 27B active, 1M context — Dr_Singularity · 2026-10-15
- StepFun open-sources Step 5 Preview: 600B MoE with 1M context, No. 2 on OpenRouter — Dr_Singularity · 2026-10-15
- Google says teams across the company are using Gemini 4 Argon on its hardest engineering problems — CurieuxExplorer · 2026-10-15
- Perplexity's new embedding models: index with a 9B model, search with 0.6B — antoine_chaffin · 2026-10-15
- 797GB of n-gram counts in CPU RAM let 1B models match 57% more compute — PangWeiKoh · 2026-10-15
- Leak: Moonshot's Kimi flagship 'x4' done with pre-training, K3.1 in internal testing — lyraxana · 2026-10-15