TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs
Paper โข 2607.17423 โข Published โข 76
How to use ZhuYuhan/TimeLens2-35B-A3B-SFT with Transformers:
# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM
processor = AutoProcessor.from_pretrained("ZhuYuhan/TimeLens2-35B-A3B-SFT")
model = AutoModelForMultimodalLM.from_pretrained("ZhuYuhan/TimeLens2-35B-A3B-SFT", device_map="auto")TimeLens2-35B-A3B-SFT is a video temporal grounding model based on Qwen/Qwen3.6-35B-A3B. Given a video and a text query, it predicts the relevant time intervals in seconds. This is the SFT version.
Paper ยท Code & Usage ยท Collection
| Benchmark | mIoU (%) |
|---|---|
| Charades-STA | 58.08 |
| ActivityNet Captions | 64.13 |
| QVHighlights | 72.73 |
| VUE-TR | 51.49 |
| VUE-TR-v2 | 50.74 |
| MomentSeeker | 29.94 |
| Ego4D-NLQ-v2 | 23.26 |
| Average | 50.06 |
The two VUE benchmarks use the available-video subsets. The average is the unweighted mean over the seven benchmarks.
For inference and evaluation, follow the official TimeLens2 repository and use ZhuYuhan/TimeLens2-35B-A3B-SFT as the model path.
Base model
Qwen/Qwen3.6-35B-A3B