MOSS-VL-Base-0408

MOSS-VL is an open vision-language model family from OpenMOSS, supporting image understanding, long-video understanding, and realtime streaming interaction.

Technical Report: https://arxiv.org/pdf/2608.15045

📌 Introduction

MOSS-VL-Base-0408 is the foundation checkpoint of the MOSS-VL series, part of the OpenMOSS ecosystem dedicated to advancing visual understanding.

Built through four stages of multimodal pretraining only, this checkpoint serves as a high-capacity offline multimodal base model. It provides strong general-purpose visual-linguistic representations across image and video inputs, and is intended primarily as the base model for downstream supervised fine-tuning, alignment, and domain adaptation.

Specifically, the pretraining pipeline is structured into the following four progressive stages:

• Stage 1: Vision-language alignment
• Stage 2: Large-scale multimodal pretraining
• Stage 3: High-quality multimodal pretraining
• Stage 4: Annealing and long-context extension

✨ Highlights

• 📐 Native Dynamic Resolution MOSS-VL-Base-0408 natively processes images and video frames at their original aspect ratios and resolutions. By preserving the raw spatial layout, it faithfully captures fine visual details across diverse formats—from high-resolution photographs and dense document scans to ultra-wide screenshots.
• 🎞️ Native Interleaved Image & Video Inputs The model accepts arbitrary combinations of images and videos within a single sequence. Through a unified end-to-end pipeline, it seamlessly handles complex mixed-modality prompts, multi-image comparisons, and interleaved visual narratives without requiring modality-specific pre-processing.

🏗 Model Architecture

MOSS-VL-Base-0408 adopts a cross-attention-based architecture that decouples visual encoding from cognitive reasoning. Natively supporting interleaved modalities, it provides a multimodal backbone for image and video understanding.

🧩 Absolute Timestamps

To help the model perceive the pacing and duration of events, MOSS-VL-Base-0408 injects absolute timestamps alongside sampled video frames, giving the reasoning process an explicit temporal reference even at the pretrained base stage.

🧬 Cross-attention RoPE (XRoPE)

MOSS-VL utilizes Cross-attention Rotary Position Embedding (XRoPE), tailored to its cross-attention-based vision-language architecture. This mechanism maps text tokens and visual features into a unified 3D coordinate space defined by Time (t), Height (h), and Width (w), improving spatial-temporal grounding during multimodal reasoning.

🚀 Quickstart

🛠️ Installation

conda create -n moss_vl python=3.12 pip -y
conda activate moss_vl
pip install -i https://pypi.org/simple --no-build-isolation -r requirements.txt

🏃 Run Inference

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

checkpoint = "path/to/checkpoint"
image_path = "data/example_image.jpg"


def load_model(checkpoint: str):
    processor = AutoProcessor.from_pretrained(
        checkpoint,
        trust_remote_code=True,
        frame_extract_num_threads=1,
    )
    model = AutoModelForCausalLM.from_pretrained(
        checkpoint,
        trust_remote_code=True,
        device_map="auto",
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
    )
    return model, processor


model, processor = load_model(checkpoint)

text = model.offline_image_generate(
    processor,
    prompt="",
    image=image_path,
    shortest_edge=4096,
    longest_edge=16777216,
    multi_image_max_pixels=201326592,
    patch_size=16,
    temporal_patch_size=1,
    merge_size=2,
    image_mean=[0.5, 0.5, 0.5],
    image_std=[0.5, 0.5, 0.5],
    max_new_tokens=256,
    temperature=1.0,
    top_k=50,
    top_p=1.0,
    repetition_penalty=1.0,
    do_sample=False,
    vision_chunked_length=64,
    use_template=False,
)

print(text)

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

checkpoint = "path/to/checkpoint"
video_path = "data/example_video.mp4"


def load_model(checkpoint: str):
    processor = AutoProcessor.from_pretrained(
        checkpoint,
        trust_remote_code=True,
        frame_extract_num_threads=1,
    )
    model = AutoModelForCausalLM.from_pretrained(
        checkpoint,
        trust_remote_code=True,
        device_map="auto",
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
    )
    return model, processor


model, processor = load_model(checkpoint)

text = model.offline_video_generate(
    processor,
    prompt="",
    video=video_path,
    shortest_edge=4096,
    longest_edge=16777216,
    video_max_pixels=201326592,
    patch_size=16,
    temporal_patch_size=1,
    merge_size=2,
    video_fps=1.0,
    min_frames=1,
    max_frames=256,
    num_extract_threads=4,
    image_mean=[0.5, 0.5, 0.5],
    image_std=[0.5, 0.5, 0.5],
    max_new_tokens=256,
    temperature=1.0,
    top_k=50,
    top_p=1.0,
    repetition_penalty=1.0,
    do_sample=False,
    vision_chunked_length=64,
    use_template=False,
)

print(text)

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

checkpoint = "path/to/checkpoint"
shared_generate_kwargs = {
    "temperature": 1.0,
    "top_k": 50,
    "top_p": 1.0,
    "max_new_tokens": 256,
    "repetition_penalty": 1.0,
    "do_sample": False,
}
shared_video_media_kwargs = {
    "min_pixels": 4096,
    "max_pixels": 16777216,
    "video_max_pixels": 201326592,
    "video_fps": 1.0,
    "min_frames": 1,
    "max_frames": 256,
}


def load_model(checkpoint: str):
    processor = AutoProcessor.from_pretrained(
        checkpoint,
        trust_remote_code=True,
        frame_extract_num_threads=1,
    )
    model = AutoModelForCausalLM.from_pretrained(
        checkpoint,
        trust_remote_code=True,
        device_map="auto",
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
    )
    return model, processor


model, processor = load_model(checkpoint)
queries = [
    {
        "images": ["data/sample_a.jpg"],
        "generate_kwargs": dict(shared_generate_kwargs),
    },
    {
        "videos": ["data/sample_b.mp4"],
        "media_kwargs": dict(shared_video_media_kwargs),
        "generate_kwargs": dict(shared_generate_kwargs),
    },
]

with torch.no_grad():
    result = model.offline_batch_generate(
        processor,
        queries,
        session_states=None,
        vision_chunked_length=64,
    )

texts = [item["text"] for item in result["results"]]

🚧 Limitations and Future Work

MOSS-VL-Base-0408 is a pretrained base checkpoint, and we are actively improving several core capabilities for future iterations:

• 📄 Stronger OCR, Especially for Long Documents — We plan to further improve text recognition, document parsing, and long-document understanding. A key focus is achieving near-lossless information extraction and understanding for extremely long and structurally complex inputs, such as accurately parsing texts, tables, and mathematical layouts from multi-page academic papers (dozens of pages) or dense PDF reports without degrading context or structural integrity.
• 🎬 Expanded Extremely Long Video Understanding — We aim to significantly extend the model's capacity for comprehending extremely long videos spanning several hours to dozens of hours. This includes advancing temporal reasoning and cross-frame event tracking for continuous analysis of full-length movies, lengthy meetings, or extended surveillance streams, enabling robust retrieval and understanding over ultra-long visual contexts.

[!NOTE] We expect future releases to continue strengthening the base model itself while also enabling stronger downstream aligned variants built on top of it.

📜 Citation

@misc{mossvl,
  title         = {MOSS-VL Technical Report},
  author        = {Wang, Pengyu and Tan, Chenkun and Zhou, Shaojun and Zhou, Qirui and Chen, Yanxin and He, Xingyang and Zeng, Huazheng and Cheng, Jijun and Wang, Chenghao and Qian, Xiaomeng and Wang, Pengfei and Huang, Zhan and Gao, Shanqing and Huang, Wei and Cao, Longjun and Ran, Wu and Liu, Jie and Zhu, Changtai and Wang, Hongkai and Tian, Yixian and Liu, Chenghao and Ye, Zhen and Wang, Xinghao and Jiang, Botian and Feng, Guoguo and Fei, Zhaoye and Li, Ruixiao and Chen, Mingshu and Gao, Yang and Cheng, Qinyuan and Li, Shimin and Qiu, Xipeng},
  year          = {2026},
  eprint        = {2608.15045},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.15045}
}

@misc{mossvideopreview,
  title         = {{MOSS-Video-Preview: Toward Real-Time Video Understanding via Cross-Attention}},
  author        = {Pengyu Wang and Chenkun Tan and Shaojun Zhou and Wei Huang and Qirui Zhou and Zhan Huang and Zhen Ye and Jijun Cheng and Xiaomeng Qian and Yanxin Chen and Xingyang He and Huazheng Zeng and Chenghao Wang and Pengfei Wang and Hongkai Wang and Shanqing Gao and Yixian Tian and Chenghao Liu and Xinghao Wang and Botian Jiang and Xipeng Qiu},
  year          = {2026},
  eprint        = {2606.07639},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.07639}
}