OpenMOSS-Team/MOSS-VL

1---2title: MOSS-VL3emoji: 🌱4colorFrom: green5colorTo: blue6sdk: gradio7sdk_version: 5.50.08python_version: "3.10"9app_file: app.py10pinned: false11short_description: 'MOSS-VL: Toward Advanced Video Understanding'12license: apache-2.013models:14  - OpenMOSS-Team/MOSS-VL-Instruct-040815tags:16  - vision-language17  - multimodal18  - image-understanding19  - video-understanding20---21 22# MOSS-VL-Instruct-0408 Demo23 24An interactive demo for **MOSS-VL-Instruct-0408**, an 11B-parameter instruction-tuned vision-language model developed by the [OpenMOSS Team](https://github.com/OpenMOSS). Built on MOSS-VL-Base-0408 through supervised fine-tuning, it serves as a high-performance offline multimodal engine with particular strength in **video understanding**.25 26## Highlights27 28- **Outstanding Video Understanding** — Long-form video comprehension, temporal reasoning, action recognition, and second-level event localization. Top-tier results on VideoMME and MLVU, with +8.3 pts on VSI-bench over Qwen3-VL-8B-Instruct.29- **Strong General Multimodal Perception** — Robust image understanding, fine-grained object recognition, OCR, and document parsing (83.9 on document/OCR benchmarks).30- **Reliable Instruction Following** — Enhanced alignment with user intent through supervised fine-tuning on diverse multimodal instruction data.31 32## Architecture33 34MOSS-VL adopts a **cross-attention-based architecture** that decouples visual encoding from cognitive reasoning:35 36- Millisecond-level latency for instantaneous responses37- Natively supports **interleaved modalities** — processes complex sequences of images and videos within a unified pipeline38- **Absolute Timestamps** injected alongside each sampled frame for precise temporal perception39- **Cross-attention RoPE (XRoPE)** — maps text tokens and video patches into a unified 3D coordinate space (time, height, width)40 41## Capabilities42 43- **Image Understanding**: scene description, object recognition, visual reasoning44- **Video Understanding**: temporal reasoning, action recognition, key event localization45- **OCR & Document Parsing**: text extraction and structured document parsing46- **Visual Question Answering**: open-ended questions about any image or video47 48## Usage49 501. Upload an **image** or **video** using the input panel, or pick one of the example prompts on the welcome screen512. Enter your question or prompt in the text box523. (Optional) Adjust generation parameters in the sidebar's **Generation Settings**534. Press **Enter** or click **Send** to get the model's response54 55> **Note**: The model weights (~22 GB) may take a few minutes to load on first use (cold start). Subsequent requests will be faster.56 57## Model Details58 59- **Model**: [OpenMOSS-Team/MOSS-VL-Instruct-0408](https://huggingface.co/OpenMOSS-Team/MOSS-VL-Instruct-0408)60- **Parameters**: 11B (BF16)61- **Base Model**: MOSS-VL-Base-040862- **License**: Apache 2.063 64## Citation65 66```bibtex67@misc{moss_vl_2026,68  title         = {{MOSS-VL Technical Report}},69  author        = {OpenMOSS Team},70  year          = {2026},71  howpublished  = {\url{https://github.com/OpenMOSS/MOSS-VL}},72  note          = {GitHub repository}73}74```75