NEWVectors or files. Pick a path.Start →
    Models/ZhuYuhan/TimeLens2-9B
    Apache-2.0

    TimeLens2-9B

    by ZhuYuhan

    A 9B TimeLens2 checkpoint on Qwen3.5 that returns the seconds where a text query happens in a video, at 50.28 average mIoU on seven benchmarks

    Identifiers
    Model ID
    ZhuYuhan/TimeLens2-9B

    Deploy TimeLens2-9B

    Single-tenant

    Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

    Overview

    TimeLens2-9B is a video temporal grounding model: give it a video and a description or question, and it returns the time intervals, in seconds, where that happens. It was published on Hugging Face on November 14, 2026 by Yuhan Zhu, first author of the TimeLens2 paper from Nanjing University's MCG lab, under Apache-2.0. It is fine-tuned from Qwen3.5-9B, where the earlier TimeLens2 checkpoints build on Qwen3-VL.

    The card reports 50.28 average mIoU over seven grounding benchmarks, above the 48.0 the paper's TimeLens2-8B reports on the same suite, with 60.56 on Charades-STA, 61.97 on ActivityNet Captions and 70.28 on QVHighlights. A 35B-A3B mixture-of-experts sibling on Qwen3.6 averages 50.39, almost the same score with a much larger total parameter count.

    Like every TimeLens2 model it is generative: it reads the clip for each query and writes the intervals as text. It suits the few clips a search returns, where you need exact start and end seconds, and is too slow to scan a library by itself.

    Architecture

    Qwen3.5-9B fine-tuned in the two TimeLens2 stages: supervised fine-tuning on TimeLens2-93K (23,793 videos, 93,232 grounding instances, with single- and multi-span answers), then GRPO with a temporal IoU reward and a matching-free temporal Wasserstein reward. The output is a set of [start, end] intervals in seconds, so one query can return several spans.

    Mixpeek SDK Integration

    # Index the library first: a Mixpeek video collection cuts each file into scenes
    # and makes them searchable. TimeLens2 then refines a found scene to exact seconds.
    import requests
    
    requests.post(
        "https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
        headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
        json={"blobs": [{"property": "video", "type": "video", "data": "s3://footage/practice-0412.mp4"}]},
    )

    Capabilities

    • Returns one or more time intervals in seconds for a text description or question about a video
    • Short and long videos, third-person and first-person (egocentric) footage
    • 50.28 average mIoU across seven grounding benchmarks, per the model card
    • Apache-2.0, so commercial use is allowed

    Use Cases on Mixpeek

    Tightening a retrieved scene to the exact seconds that match the query
    Cutting highlight clips from a text request
    Labelling when an action happens in review, training or sports footage
    Answering "when does X happen" questions over a found video

    Benchmarks

    DatasetMetricScoreSource
    Seven temporal grounding benchmarks (unweighted average)mIoU50.28Model card: ZhuYuhan/TimeLens2-9B (self-reported; TimeLens2-8B: 48.0)
    Charades-STAmIoU60.56Model card (self-reported)
    ActivityNet CaptionsmIoU61.97Model card (self-reported)
    QVHighlightsmIoU70.28Model card (self-reported)
    VUE-TR (available-video subset)mIoU54.04Model card (self-reported)
    MomentSeekermIoU30.23Model card (self-reported)
    Ego4D-NLQ v2mIoU22.69Model card (self-reported)

    Performance

    Input SizeA video and a text query; frames are sampled as in the TimeLens2 repository's inference code
    Embedding Dimn/a (outputs time intervals as text)
    GPU LatencyInput dependent
    GPU ThroughputBatch dependent
    GPU MemoryModel dependent

    A generative model that reads the clip for every query, so run it on the few clips a retriever returns rather than across a library. Long first-person video stays hard: Ego4D-NLQ v2 is 22.69 mIoU against 70.28 on QVHighlights. We have not measured it.

    Frequently Asked Questions

    What is TimeLens2-9B?

    A video temporal grounding model fine-tuned from Qwen3.5-9B. Given a video and a text query, it returns the time intervals in seconds where the query is true. It was published on November 14, 2026 under Apache-2.0 by the first author of the TimeLens2 paper.

    Is TimeLens2-9B better than TimeLens2-8B?

    On the reported numbers, yes: 50.28 average mIoU over seven benchmarks against 48.0 for the 8B checkpoint. The 9B is built on Qwen3.5-9B and the 8B on Qwen3-VL-8B. Both scores are self-reported.

    What is video temporal grounding?

    Finding when something happens in a video: the input is a video and a description, the output is the start and end seconds of each matching moment. It is scored with mIoU, the average overlap between predicted and true intervals.

    Can TimeLens2-9B search a whole video library?

    Not on its own. It reads the clip for each query, so it suits the handful of clips a search returns. Index the library with embedding-based search first, then run TimeLens2 on the top results for exact seconds.

    Can I use TimeLens2-9B commercially?

    Yes. It is released under Apache-2.0.

    How do I use TimeLens2-9B with Mixpeek?

    Index the footage in a Mixpeek video collection and search it with a retriever, which returns scenes with timestamps. Run TimeLens2-9B on the top scene with the same query when you need tighter start and end seconds than the scene boundaries.

    Specification

    OrganizationZhuYuhan
    Retriever-
    Parameters9.4B
    LicenseApache-2.0
    Downloads/moN/A
    Likes2

    Research Paper

    TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs (Zhu et al., July 2026)

    arxiv.org

    Build a pipeline with TimeLens2-9B

    Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

    Run it on your own data, free