Tarsier-2

Tarsier-2 (arXiv 2501.07888, January 2025) is ByteDance Research's second-generation video description model, and it is unapologetically optimized for one thing: writing detailed, accurate descriptions of what happens in a video. Three upgrades over Tarsier-1: pretraining data scaled from 11M to 40M video-text pairs, fine-grained temporal alignment during supervised fine-tuning, and DPO on model-sampled preference data. The authors report Tarsier2-7B beating GPT-4o by 2.8% F1 and Gemini-1.5-Pro by 5.8% F1 on DREAM-1K — the 1,000-clip description benchmark the same team created — and winning human side-by-side comparisons against GPT-4o by 8.6%. Beyond description, they claim new state-of-the-art 7B results on 15 public benchmarks spanning video QA, grounding and hallucination tests. The repo (bytedance/tarsier) is Apache 2.0 with public weights, and the DREAM-1K annotations plus the AutoDQ evaluation method are released with it. No hosted API is offered.

• DREAM-1K detailed description: +2.8% F1 over GPT-4o, +5.8% over Gemini-1.5-Pro (author-reported)\n- Human side-by-side: +8.6% over GPT-4o, +24.9% over Gemini-1.5-Pro (author-reported)\n- Claimed 7B SOTA on 15 public benchmarks (MVBench, Video-MME, LongVideoBench, MLVU, TemporalBench and others; author-reported)\n- No independent replication of the headline figures found