---
title: Leum-VL-8B-preview0320
canonical_url: "https://www.modelscope.cn/models/leum-team/Leum-VL-8B-preview0320"
md_url: "https://www.modelscope.cn/models/leum-team/Leum-VL-8B-preview0320.md"
repository: leum-team/Leum-VL-8B-preview0320
chinese_name: Leum-VL-8B-preview0320
last_updated: 2026-03-24
license: mit
model_type:
  - qwen3_vl
architectures:
  - Qwen3VLForConditionalGeneration
base_model_relation: finetune
parameters: 8.8B
tensor_type:
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - Pytorch
language:
  - en
  - zh
downloads: 18
stars: 0
---

# Leum-VL-8B-preview0320

> Leum-VL-8B-preview0320 - leum-team 在 ModelScope 开源的模型。Leum-VL-8B-preview0320

leum-team/Leum-VL-8B-preview0320 是 ModelScope 魔搭社区上的 8.8B 参数机器学习模型，采用 mit 许可。

- **Repository**: leum-team/Leum-VL-8B-preview0320
- **License**: mit
- **Parameters**: 8.8B
- **Downloads**: 18
- **Stars**: 0
- **Last updated**: 2026-03-24

Source: https://www.modelscope.cn/models/leum-team/Leum-VL-8B-preview0320

---

<p align="center">
  <img src="./assets/leum_logo.png" alt="Leum-VL logo" width="650" />
</p>

# Leum-VL-8B-preview0320

## 面向互联网原生短视频的时间轴结构理解 / Timeline-Grounded Structural Understanding for Internet-Native Short Video

[技术报告 Technical Report (arXiv)](https://arxiv.org/abs/2603.20354) | [PDF](./assets/report.pdf) | [Hugging Face](https://huggingface.co/leum-team/Leum-VL-8B-preview0320) | [ModelScope](https://modelscope.cn/models/leum-team/Leum-VL-8B-preview0320) | [GitHub](https://github.com/leum-team/leum-vl) | [联系 Contact](mailto:chaiming@sihe.ai)

Leum-VL-8B 是一个 8B 开源权重视频语言模型，面向互联网原生短视频的时间轴结构理解。它将短视频解析为与时间轴对齐、可追溯到证据、可供机器使用的结构化输出，覆盖 subject、aesthetics、camera language、editing、narrative 和 observable dissemination strategy 六个维度。

Leum-VL-8B is an 8B open-weight video-language model for timeline-grounded structural understanding of internet-native short video. It parses short videos into timeline-aligned, evidence-linked, machine-usable structure across six dimensions: subject, aesthetics, camera language, editing, narrative, and observable dissemination strategy.

在本 README 中，observable dissemination strategy 指在提供相关上下文时，模型可利用视频周边的可观测传播线索，例如 title、hashtags、cover 和 comments，对 retention-oriented design、platform-native packaging 和 engagement context 做结构化分析。

In this model card, observable dissemination strategy refers to retention-oriented design, platform-native packaging, and engagement context around the video when such context is provided, including title, hashtags, cover, and comments.

<p align="center">
  <img src="./assets/benchmark_comparison.png" alt="Benchmark comparison" width="100%" />
</p>

## 更新 / Updates

- `[2026-03-24]` ModelScope README 与资源文件已更新到当前技术报告版本，并与 Hugging Face model card 对齐。
- `[TBD]` 推理 API 示例与部署命令将在工程验证完成后补充。
- `[TBD]` 正式的 YAML / JSON schema 与任务提示模板将在后续版本补充。
- `[2026-03-24]` The ModelScope README and assets have been refreshed to align with the current technical report and the Hugging Face model card.
- `[TBD]` Inference API examples and deployment commands will be added after engineering validation.
- `[TBD]` The formal YAML / JSON schema and task-specific prompt templates will be added in a later update.

## 特点 / Highlights

- 理解的不只是场景内容，也包括剪辑与结构。
- 面向互联网原生短视频结构解析而设计。
- 输出为时间轴对齐、可供机器消费的结构化结果。
- 采用 vision encoder、projector 与 language model 全参数训练。
- Understands edits, not just scene content.
- Built for internet-native short-video structure.
- Produces timeline-grounded, machine-usable outputs.
- Trained with full-parameter updates across the vision encoder, projector, and language model.

## Benchmark Snapshot / Benchmark Snapshot

下表与当前技术报告中的 benchmark 表保持一致。  
The table below mirrors the current benchmark table in the technical report.

| Category | Benchmark | Leum-VL-8B | Qwen3-VL-8B<sup>1</sup> | Keye-VL-8B Thinking<sup>2</sup> | GLM-4.1V-9B Thinking<sup>3</sup> | MiniCPM-V-4.5-8B<sup>4</sup> |
| --- | --- | ---: | ---: | ---: | ---: | ---: |
| General VQA | MMBench-EN test | **84.8** | 84.5 | 92.0 | 85.8 | 84.2 |
|  | MMBench-CN test | 83.9 | **84.7** | - | 84.7 | - |
|  | HallusionBench | 56.5 | 61.1 | 62.7 | **63.2** | 61.2 |
|  | RealWorldQA | **73.2** | 71.5 | 73.5 | - | 72.1 |
|  | MMStar | 67.5 | 70.9 | **80.5** | 72.9 | 72.1 |
|  | BLINK | 65.2 | **69.1** | 54.9‡ | 65.1 | 42.0‡ |
| Document & OCR | OCRBench | 85.4 | **89.6** | 86.6 | 84.2 | 89.0 |
|  | DocVQA test | 95.7 | **96.1** | 93.4‡ | 93.3‡ | 94.7 |
|  | TextVQA val | **85.0** | 82.8‡ | 81.5‡ | 79.6‡ | 82.2 |
|  | ChartQA test | 85.3 | **89.6** | 94.1‡ | 70.0‡ | 87.4 |
| Video Understanding | Video-MME w/o sub. | 70.8 | **71.4** | 73.0 | 68.2 | 67.9 |
|  | MVBench | **70.0** | 68.7 | 56.9‡ | 68.4 | 60.5‡ |
|  | TempCompass | **74.3** | 74.3‡ | **75.5** | 72.3‡ | 72.7‡ |
|  | MotionBench | **61.6** | 56.9‡ | 55.1‡ | 59.0 | 59.7 |
|  | FAVOR-Bench | **58.9** | 54.1 | - | - | 56.0 |
|  | LongVideoBench | 64.6 | 62.4‡ | **66.0** | 65.7‡ | 63.9 |
|  | Tomato | **36.7** | 35.7‡ | 33.0‡ | 30.0‡ | 29.8‡ |
|  | Charades-STA mIoU | **59.4** | 56.0 | - | - | - |

<sup>1</sup> Qwen3-VL report.  
<sup>2</sup> Keye-VL report.  
<sup>3</sup> GLM-4.1V report.  
<sup>4</sup> MiniCPM-V report.  
`‡` Reported values reproduced from referenced public reports as cited in the technical report.

评测设置 / Eval setting: `vLLM v0.17.1-cu130`, `FPS=4`, `max 768 frames`, `max 50K tokens/video`.

我们同时构建了 FeedBench，用于结构敏感型短视频理解评测。`[TBD]` 公开 benchmark 页面和数据集链接将在后续补充。  
We also construct FeedBench, a benchmark for structure-sensitive short-video understanding. `[TBD]` The public benchmark page and dataset link will be added after release.

## 差异化能力 / What Makes It Different

传统视频语言模型通常能够描述场景、回答事件类问题或读取画面文字，但往往不擅长解释为什么发生某个 cut、某段内容承担什么 narrative 作用、短视频使用了哪些 retention 策略，或者视频内容如何与平台原生 packaging 和 audience-response context 对齐。Leum-VL-8B 将视频理解建模为 timeline-grounded structural parsing，并可在提供上下文时结合平台侧线索做分析。

Conventional video-language models can often describe scenes, answer event-centric questions, or read on-screen text. They are typically less reliable at explaining why a cut happens, what narrative role a segment serves, what retention tactics a short video uses, or how video content aligns with platform-native packaging and audience-response context. Leum-VL-8B addresses this gap by treating video understanding as timeline-grounded structural parsing that can incorporate linked platform context.

| Paradigm | Can capture | Often misses |
| --- | --- | --- |
| Dense captioning | Scene and event semantics | Cut rationale, hook design, narrative role |
| VQA | Event-centric reasoning | Timeline-level structural grammar |
| OCR / subtitle parsing | On-screen text | How text interacts with pacing and packaging |
| Shot boundary detection | Physical cuts | Why the cut matters |
| Leum-VL-8B (SV6D) | Multi-layer timeline-grounded structure | - |

## SV6D 表示 / SV6D Representation

SV6D 将视频拆解为六个互补维度，每个维度都与时间轴对齐，并可以落到可观测证据上。  
SV6D decomposes video into six complementary dimensions, each grounded on the timeline and tied to observable evidence.

| Dimension | What it captures | Observable evidence or linked context | Downstream value |
| --- | --- | --- | --- |
| Subject | Who or what is present and how it changes | People, objects, actions, scene changes | Retrieval, indexing |
| Aesthetics | Visual style and perceptual tone | Color, lighting, composition, overlay style | Style analysis, generation control |
| Camera language | Shot size, angle, and movement | Close-up, pan, zoom, handheld, tracking | Cinematography parsing |
| Editing | Transitions and pacing logic | Hard cuts, jump cuts, rhythm changes, tension release | Edit analysis, editing assistance |
| Narrative | Timeline-grounded function of segments | Hook, setup, progression, reveal, payoff | Story analysis, clip structuring |
| Observable dissemination strategy | Retention-oriented design, packaging, and engagement context around the video | Topic framing, first-3-second hook design, subtitle emphasis, cover-title-hashtag consistency, video-comment alignment | Retention analysis, packaging analysis, comment analysis, creator tooling |

Leum-VL-8B 通过 continual pretraining、SFT 和 RLHF alignment 的全参数训练实现 SV6D 目标。模型从多模态 instruct checkpoint 初始化，训练目标是输出 temporally aligned、evidence-linked、machine-usable 的短视频结构解析结果。

Leum-VL-8B realizes the SV6D objective through full-parameter training spanning continual pretraining, supervised fine-tuning, and RLHF alignment. Initialized from a multimodal instruct checkpoint, it is trained to produce temporally aligned, evidence-linked, machine-usable outputs for short-video structural parsing.

## 模型概览 / Model Snapshot

| Item | Value |
| --- | --- |
| Model name | Leum-VL-8B |
| Model type | Video-language model |
| Parameters | 8B |
| Base model | [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) |
| Training scope | Full-parameter training over vision encoder, projector, and language model |
| Training stages | Continual pretraining -> SFT -> RLHF alignment |
| Primary task | Timeline-grounded structural parsing for short video |
| Output | Structured YAML or text reports aligned to the timeline |
| Release type | Open weights |
| License | MIT |

## 输出格式 / Output Format

Leum-VL-8B 的输出设计为可直接被下游系统消费。下面示例仅用于展示输出风格。`[TBD]` 字段名与 schema 在正式发布前仍可能调整。  
Leum-VL-8B is designed to produce outputs that are directly usable by downstream systems. The following is an illustrative example. `[TBD]` Field names and schema are not final until the official release schema is published.

```yaml
timeline:
  - span: "<00:00.0-00:07.0>"
    content_structure: "Opening Hook"
    structure_description: |
      At the center of the frame stands a multi-tiered pagoda-style structure against the night sky. The eaves on each level are lined with alternating cyan and golden light strips. The tower continuously emits golden linear fireworks outward, forming parabolic trajectories that spread to both sides. Gradually, red light beams begin to dominate, and the atmosphere of the fireworks shifts from tranquil to intense, completing the emotional setup of the opening.
    shots:
      - span: "<00:00.0-00:07.0>"
        subject_analysis: |
          A multi-tiered pagoda-style building stands at the center of the frame, tapering upward level by level. Each tier is decorated with surrounding light strips that alternate between cyan and golden illumination, with a pointed ornament at the top. From each level of the structure, golden linear fireworks continuously shoot outward, forming parabolic trajectories that spread to the sides and upward. As the scene progresses, denser red light beams shoot from the tower, and the lighting color of the structure also changes, with red gradually becoming dominant. The background is a black night sky. At the bottom of the frame, there is a horizontal railing, beyond which a large crowd appears as black silhouettes. Among them are dense rectangular lights from handheld electronic device screens.
        shot_breakdown:
          shot_size: "Wide shot"
          camera_position: "Low angle position"
          camera_angle: "Upward angle"
          focal_length: "Wide-angle"
          camera_movement: "Handheld shake"
          depth_of_field: "Deep depth of field"
        aesthetic_analysis:
          light_source_type: "Artificial light"
          lighting_direction: "Backlight / rim light"
          light_hardness: "Hard light"
          contrast: "High contrast"
          saturation: "High"
          color_temperature_tone: "Mixed warm and cool tones"
          base_tone: "Low-key"
          composition: "Symmetrical composition"
    emotional_curve:
      - time: "<00:00.0>"
        emotion_level: 0 (Calm)
        description: |
          The video begins with the Wunüzhou pagoda standing under the night sky, presenting a gorgeous and tranquil scene as a buildup.
      - time: "<00:02.1>"
        emotion_level: 1 (Engaged)
        description: |
          The first wave of fireworks bursts on both sides of the tower, breaking the stillness with visual motion and raising emotional engagement.

  - span: "<00:07.0-00:14.9>"
    content_structure: "Core Content"
    structure_description: |
      The fireworks display enters its visual climax phase. Slender white firework streaks and dense colorful bursts appear alternately. Thick smoke spreads, gradually obscuring the tower's outline. The brightness and colors shift dramatically from white to emerald green to pink. Strong light penetrates the smoke, forming a hazy glow that delivers continuous visual impact and immersion, sustaining the emotional peak through the ending.
    shots:
      - span: "<00:07.0-00:09.5>"
        subject_analysis: |
          The multi-tiered pagoda stands at the center of the frame. The lights on each level form red circular rings, with blue light visible in the gaps. The building's outline is faintly visible through the smoke. Slender bright white firework streaks continuously shoot diagonally upward from the tower, leaving straight trails in the air. Dense gray-white smoke forms around the structure and rises upward, partially obscuring architectural details. At the bottom of the frame, a bridge railing emits a cyan-blue glow, with silhouettes of spectators holding recording devices in front of it. The surrounding background is a dark night sky, showing slight gray variations under the illumination of fireworks.
        shot_breakdown:
          shot_size: "Wide shot"
          camera_position: "Low angle position"
          camera_angle: "Upward angle"
          focal_length: "Wide-angle"
          camera_movement: "Handheld shake"
          depth_of_field: "Deep depth of field"
        aesthetic_analysis:
          light_source_type: "Artificial light"
          lighting_direction: "Backlight / rim light"
          light_hardness: "Hard light"
          contrast: "High contrast"
          saturation: "High"
          color_temperature_tone: "Mixed warm and cool tones"
          base_tone: "Low-key"
          composition: "Symmetrical composition"
      - span: "<00:09.5-00:14.9>"
        subject_analysis: |
          The lower part of the frame shows the top structure of a building emitting bright golden light, with a square, block-like appearance and a smooth surface. Dense firework streaks spray upward from the top in a vast fan shape. The fireworks initially appear bright white, then shift to emerald green, and finally turn pink. As the fireworks erupt, large amounts of smoke spread outward, and light passing through the smoke creates a soft halo effect. The black night sky is illuminated by the fireworks. Along the bottom edge of the frame, silhouettes of raised rectangular objects occasionally appear in extremely low light conditions.
        shot_breakdown:
          shot_size: "Wide shot"
          camera_position: "Low angle position"
          camera_angle: "Upward angle"
          focal_length: "Telephoto"
          camera_movement: "Handheld shake"
          depth_of_field: "Medium depth of field"
        aesthetic_analysis:
          light_source_type: "Artificial light"
          lighting_direction: "Backlight / rim light"
          light_hardness: "Hard light"
          contrast: "High contrast"
          saturation: "Medium"
          color_temperature_tone: "Warm"
          base_tone: "Mid-tone"
          composition: "No obvious compositional intent"
    emotional_curve:
      - time: "<00:08.1>"
        emotion_level: 2 (Climax)
        description: |
          Fireworks of varying forms and colors alternate continuously, maintaining a high density of visual output and immersing the audience in a sense of exhilaration.
```

根据任务模式不同，observable dissemination strategy 可以表现为 segment-level retention cues，也可以表现为 clip-level packaging 或 comment alignment 输出。  
Depending on the task mode, observable dissemination strategy may be expressed either as segment-level retention cues or as clip-level packaging or comment alignment outputs.

## 快速开始 / Get Started

- 主要输入：MP4。
- 可选上下文：title、hashtags、cover images、comments。
- 推荐部署框架：vLLM。
- 规划中的任务模式：`sv6d_parse`、`summary`、`edit_suggestions`、`retention_analysis`、`comment_analysis`、`packaging_alignment`。
- Primary input: MP4.
- Optional associated context for supported task modes: title, hashtags, cover images, and comments.
- Recommended deployment framework: vLLM.
- Planned task modes: `sv6d_parse`, `summary`, `edit_suggestions`, `retention_analysis`, `comment_analysis`, `packaging_alignment`.

`[TBD]` 官方推理示例、部署配方与验证后的 schema 将在后续补充。  
`[TBD]` Official inference examples, deployment recipes, and the validated schema will be added later.

## 使用场景 / Use Cases

- 短视频结构解析：识别 hook、setup、progression、reveal、payoff。
- 剪辑分析：理解 cut、节奏变化与 shot-level transitions。
- retention 分析：识别 topic promise、first-3-second hook design、curiosity gaps、multi-hook chaining 与 payoff timing。
- 字幕密集型互联网视频理解：将 overlays、stickers、UI-like layouts 作为结构信号。
- packaging alignment：评估 cover、title、hashtags 与视频内容、结构是否一致。
- comments 分析：结合 comments 做视频对齐与 comment-aware summarization。
- 检索与索引：按结构模式而不仅仅是对象或事件检索视频。
- 创作者工具：支持 edit review、packaging review 与最小修改建议。
- Short-video structural parsing: identify hooks, setup, progression, reveal, and payoff.
- Edit analysis: reason about cuts, pacing changes, and shot-level transitions.
- Retention analysis: identify topic promise, first-3-second hook design, curiosity gaps, multi-hook chaining, and payoff timing.
- Subtitle-heavy internet video understanding: analyze overlays, stickers, and UI-like layouts as structural signals.
- Packaging alignment: assess whether cover image, title, and hashtags match the video's actual content and structure.
- Comment analysis: align videos with associated comments and support comment-aware summarization.
- Retrieval and indexing: search videos by structural patterns rather than only objects or events.
- Creator tooling: support edit review, packaging review, and minimal revision suggestions.

## 局限性 / Limitations

- observable dissemination strategy 指的是视频周边可观测到的 retention、packaging 与 engagement 相关信号，不代表对 virality、reach、CTR 或平台分发结果的因果预测。
- comments 相关输出依赖 comments 的可用性、时效性与平台特性。
- cover-title-hashtag alignment 反映的是语义与策略一致性，不保证真实效果。
- 在损坏视频、极端视觉杂讯、OCR 可读性很差或训练分布之外的编辑风格上，性能可能下降。
- 时间边界与 segment labels 是近似的，不是 frame-perfect。
- narrative 与 dissemination 标签具有解释性，可能随文化、语言、平台和标注者而变化。
- 结构化输出应作为辅助分析使用，不应单独用于高风险决策。
- Observable dissemination strategy refers to observable retention, packaging, and engagement-related context signals around the video, not causal prediction of virality, reach, CTR, or platform distribution outcomes.
- Comment-related outputs depend on the availability, freshness, and platform specificity of associated comments.
- Cover-title-hashtag alignment reflects semantic and strategic consistency, not guaranteed performance.
- Performance may degrade on corrupted videos, visually cluttered inputs, poor OCR readability, or editing conventions far outside the training distribution.
- Timestamp boundaries and segment labels are approximate rather than frame-perfect.
- Narrative and dissemination labels are interpretive and may vary across cultures, languages, platforms, or annotators.
- Structural outputs should be used as assistive analysis, not as the sole decision-maker in high-stakes settings.

## 非目标用途 / Out-of-Scope Use

- 未经外部验证的内容效果因果判断。
- 完全自动化的 moderation 或 enforcement 决策。
- 缺少人工复核的高风险判断。
- Causal claims about content performance without external validation.
- Fully automated moderation or enforcement decisions.
- High-stakes judgment without human review.

## 引用 / Citation

```bibtex
@article{LeumVL,
  title={Leum-VL Technical Report},
  author={Yuxuan He and Chaiming Huang and Yifan Wu and Hongjun Wang and Chenkui Shen and Jifan Zhang and Long Li},
  journal={arXiv preprint arXiv:2603.20354},
  year={2026}
}
```

## 许可证 / License

本模型以 **MIT License** 发布。  
This model is released under the **MIT License**.
