GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
XinNUS
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
XinNUS
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Temporal grounding: locate the time interval matching a textual event description in a source video. This is the model used for the OMTG + Ours TimeLens comparison. The main DOVE captioning model is instead Generalized-CycleGRPO-Temporal-4B. Do not interchange these two releases when reproducing the paper tables.
| TimeLens dataset | R1@0.3 | R1@0.5 | R1@0.7 | tIoU |
|---|---|---|---|---|
| Charades-STA | 72.9 | 60.2 | 32.7 | 51.8 |
| ActivityNet Captions | 65.4 | 55.5 | 37.6 | 49.0 |
| QVHighlights | 80.4 | 70.9 | 53.7 | 64.5 |
All values are percentages under the paper's TimeLens protocol.
The repository is a standard Transformers Qwen3-VL checkpoint. Its original float32 weight serialization is retained. You can select bfloat16 for GPU inference without downloading a separate model variant.
python
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessor model_id = "XinNUS/Generalized-CycleGRPO-Temporal-OMTG-4B"model = AutoModelForImageTextToText.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto").eval()processor = AutoProcessor.from_pretrained(model_id)Provide the complete source video and the textual event query using the processor's multimodal chat template. For example, ask for the start and end times of the person opening a door. Use the official OMTG/TimeLens prompt, video-sampling, and parsing protocol for the reported numbers. A generic chat prompt is only an interface demonstration, not benchmark reproduction.
The video implementation and task-routing utilities are available in
generalized/.
No SAMTok mask tokenizer is required for temporal interval outputs.
Ambiguous queries and repeated or visually similar occurrences can produce incorrect intervals. This release is intended for research and is not a validated event-detection system for safety-critical use. It builds on OMTG and the Qwen3-VL architecture/tokenizer. See NOTICE and LICENSE for attribution and license terms.
Xin Zhang, Haochen Wang, Beibei Lin, Zhaoxiang Zhang, Robby T. Tan, Jason Li, and Dacheng Tao.