TimePLE-8B
TimePLE-8B is a Qwen3-VL-8B-Instruct model fine-tuned for video temporal grounding. Given a video and a natural-language event description, it represents and predicts the event's start/end span through a duration-adaptive TimePLE codec.
This is an inference release: it contains the selected stage-2 model weights and tokenizer/processor assets. The TimePLE model implementation is maintained in the public TimePLE source repository instead of being duplicated in this model repository. Optimizer, scheduler, RNG, and distributed-training states are intentionally excluded.
Requirements and loading
Install the TimePLE package from its public GitHub repository first, then install the pinned runtime dependencies from this directory:
pip install "timeple @ git+https://github.com/KlingAIResearch/TimePLE.git"
pip install -r requirements.txt
Importing timeple registers the TimePLE configuration, model, and processor with Transformers. The model repository contains no executable Python code and does not require trust_remote_code=True.
import torch
import timeple # registers TimePLE with the Transformers AutoClasses
from transformers import AutoModelForImageTextToText, AutoProcessor
# A downloaded local model directory also works.
model_id = "KlingTeam/TimePLE"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id,
dtype=torch.bfloat16,
device_map="auto",
).eval()
The TimePLE source repository provides the video preparation, temporal decoding, and benchmark runners. For example:
SUITE=charades_sta bash scripts/eval/run_suite.sh --models timeple_8b
The plain Transformers load above verifies and exposes the model; use the repository's TimePLE inference/evaluation path for decoded temporal intervals because it supplies video durations and the task-specific post-processing contract.
Model details
- Base model: Qwen/Qwen3-VL-8B-Instruct
- Architecture:
Qwen3VLForConditionalGenerationWithTimePLECodec - Task: video temporal grounding / moment retrieval
- Weight dtype: bfloat16
- Serialization: four sharded safetensors files
- Vision tower: frozen during stage-2 SFT
- Language model, TimePLE codec, and MLP interface adapter: trainable during stage-2 SFT
- License: Apache-2.0
The tokenizer vocabulary has 151,671 entries. It preserves all 13 Qwen3-VL additional special tokens and appends:
| Token | ID | Purpose |
|---|---|---|
| `< | TIMESTAMP | >` |
| `< | TIMESPAN | >` |
The TimePLE codec uses a 128×128 canonical span grid, a duration-adaptive residual decoder, and a learned MLP interface adapter to the 4096-dimensional language-model representation.
Intended use
This model is intended for research and evaluation of temporal localization in videos from a textual event query. Suitable uses include moment retrieval experiments, temporal-grounding ablations, and further task-specific fine-tuning.
Limitations
- Temporal quality depends on frame sampling, video duration metadata, and the inference pipeline's video preprocessing.
- The training data and prompts are primarily English; other languages and unseen domains have not been established here.
- Long, ambiguous, repeated, or extremely brief events may yield unstable boundaries.
- The reported validation result comes from the training-time validation mixture and should not be compared directly with benchmark test results produced under different sampling or evaluation protocols.
- The model inherits limitations and potential biases of Qwen3-VL and of the stage-1/stage-2 training data.
Release contents and integrity
Model parameters—including all TimePLE codec and adapter parameters—are stored in the indexed safetensors shards. No pickle-based .bin, .pt, or .pth weight file is required. The model directory contains no executable Python implementation; AutoClass registration is provided by the separately installed timeple package.
See LICENSE for the release license. Users must also comply with the terms applicable to the base model, input videos, and any datasets they use.
Citation
Paper: TimePLE: Rethinking Temporal Representation for Video Temporal Grounding