---
license: apache-2.0
language:
- en
- zh
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- wemm
- multimodal
- video
- audio
---
# WeMM-Video-9B
[](https://huggingface.co/collections/tencent/wemm-video) [](https://arxiv.org/abs/2611.26409) [](https://github.com/Tencent/WeMM-Video) [](https://huggingface.co/papers/2611.26409)
## Overview
**WeMM-Video-9B** is the dense 9B checkpoint in the WeMM-Video family. WeMM-Video is a Chinese鈥揈nglish bilingual family for joint audio-visual understanding and reasoning. It accepts text, images, audio, and video and generates text responses.
- **Weights** 路 BF16 Safetensors 路 19.54 GiB of weight files
- **Quantized variants** 路 [FP8 Block128](https://huggingface.co/tencent/WeMM-Video-9B-FP8) 路 [AWQ W4A16](https://huggingface.co/tencent/WeMM-Video-9B-AWQ) 路 [GGUF Q4_K_M](https://huggingface.co/tencent/WeMM-Video-9B-GGUF) 路 [NVFP4 W4A16](https://huggingface.co/tencent/WeMM-Video-9B-NVFP4)
### Capabilities
WeMM-Video can combine video frames with the original soundtrack to describe events, transcribe speech, and reason about what is seen and heard. It also accepts silent video, standalone audio, images, and text, with responses in Chinese or English.
### Architecture and training
Vision and audio encoders turn media into tokens that the language model processes alongside text. Video frames carry timestamps. The family was pretrained on approximately 1.2 trillion multimodal tokens, then post-trained with multi-teacher OPD, anti-repetition MrDPO, and specialist reinforcement learning.
See the [project overview](https://github.com/Tencent/WeMM-Video#overview) for family details.
## Quick Start
Set up the [Transformers environment](https://github.com/Tencent/WeMM-Video#install-dependencies). From the WeMM-Video repository root, run the example on an MP4 with a soundtrack. The script loads the model and processor from the Hub and processes both the frames and the original audio track.
```bash
python examples/transformers_video.py --model tencent/WeMM-Video-9B \
--video ./clip.mp4 --prompt "Describe what happens and summarize what you hear."
```
Use `--revision` to pin the model and processor to a Hub commit. Size GPU memory for the full weight files, media inputs, and generation cache.
### Input examples in Python
Run from the project root to use the bundled media. Each call sends one request. For video, `video` processes frames alone and `video_audio` includes the original soundtrack. Both take the file path in the `video` field.
```python
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "tencent/WeMM-Video-9B"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
model_id, trust_remote_code=True, dtype=torch.bfloat16, device_map="auto",
).eval()
def ask(prompt, *media):
content = [*media, {"type": "text", "text": prompt}]
inputs = processor.apply_chat_template(
[{"role": "user", "content": content}],
tokenize=True, return_dict=True, return_tensors="pt",
add_generation_prompt=True, enable_thinking=False,
).to(device=model.device, dtype=model.dtype)
with torch.inference_mode():
output = model.generate(**inputs, max_new_tokens=128, do_sample=False)
answer_ids = output[0, inputs.input_ids.shape[1]:]
return processor.tokenizer.decode(answer_ids, skip_special_tokens=True)
print(ask("A 20-minute meeting starts at 09:45. When does it end?"))
print(ask("Describe the planet in this image.", {
"type": "image", "image": "examples/media/apollo-11-earth.jpg",
}))
print(ask("Transcribe the English speech in this recording.", {
"type": "audio", "audio": "examples/media/apollo-11-small-step.wav",
}))
clip = dict(
video="examples/media/apollo-11-first-steps.mp4",
fps=1.0, min_frames=4, max_frames=16, total_pixels=262144,
)
print(ask("Describe what happens on screen.", {"type": "video", **clip}))
print(ask("Describe the scene and summarize the speech.", {"type": "video_audio", **clip}))
```
## Chat Modes
Instruct mode is the default. When calling `processor.apply_chat_template`, pass `enable_thinking=False` for a direct response or `enable_thinking=True` for a reasoning response.
## Serving
Install the [matching WeMM runtime patch](https://github.com/Tencent/WeMM-Video/blob/main/patch/README.md#manual-install-and-patch). The [serving guide](https://github.com/Tencent/WeMM-Video/blob/main/docs/serving.md#runtime-profiles) lists the dependency profiles and Docker images. Choose one server below.
### vLLM
Use patched vLLM 0.18.0 for the command below. The release guide also covers patched vLLM 0.31.0 for BF16. See the [vLLM recipe](https://github.com/Tencent/WeMM-Video/blob/main/docs/serving.md#serve-with-vllm) for setup details.
```bash
MODEL=tencent/WeMM-Video-9B
vllm serve "$MODEL" --served-model-name "$MODEL" --trust-remote-code --dtype bfloat16 \
--host 127.0.0.1 --port 8000 --max-model-len 8192 \
--no-enable-prefix-caching --mamba-ssm-cache-dtype float32 \
--limit-mm-per-prompt '{"image":4,"video":2,"audio":2}' \
--media-io-kwargs '{"video":{"fps":2,"min_frames":4,"max_frames":16,"total_pixels":262144}}'
```
### SGLang
Use the [WeMM BF16 recipe for patched SGLang 0.5.16 or 0.5.19](https://github.com/Tencent/WeMM-Video/blob/main/docs/serving.md#serve-with-sglang).
```bash
MODEL=tencent/WeMM-Video-9B
python -m sglang.launch_server --model-path "$MODEL" --served-model-name "$MODEL" --trust-remote-code \
--host 127.0.0.1 --port 8000 --dtype bfloat16 --context-length 8192 \
--attention-backend triton --linear-attn-backend triton \
--limit-mm-data-per-request '{"image":4,"video":2,"audio":2}'
```
### Input examples for the serving API
With the server running on port 8000, use the bundled request client from the project root in a second terminal. It sends media inline to the OpenAI-compatible chat endpoint.
```bash
MODEL=tencent/WeMM-Video-9B
python examples/chat.py text --model "$MODEL"
python examples/chat.py image --model "$MODEL"
python examples/chat.py audio --model "$MODEL"
python examples/chat.py video --model "$MODEL"
python examples/chat.py video-audio --model "$MODEL"
```
| Input | Request content |
| --- | --- |
| Text | Text string in `messages[].content` |
| Image | `image_url` with `image_url.url` |
| Audio | `audio_url` with `audio_url.url` |
| Video frames | `video_url` with `video_url.use_audio` set to `false` |
| Video with soundtrack | `video_url` with `video_url.use_audio` set to `true` |
Set `fps` and `max_frames` inside `video_url` to control sampling for a clip. Set `chat_template_kwargs.enable_thinking` to `true` for Thinking mode. See the [complete video request](https://github.com/Tencent/WeMM-Video/blob/main/docs/serving.md#send-a-video-request).
## Validation
Patched vLLM loaded the checkpoint on H20 and completed generation checks for text, image, audio, and video with a soundtrack. The local checkpoint also passed direct Transformers loading and a CPU text generation check.
## Evaluation
See the [WeMM-Video-35B-A3B evaluation results](https://huggingface.co/tencent/WeMM-Video-35BA3B#evaluation) for the benchmark comparison in the technical report.
## License
Apache License 2.0. See the [included license](LICENSE) and the [collection license and third-party attribution](https://github.com/Tencent/WeMM-Video/blob/main/LICENSE). Third-party components remain subject to their original terms.
## Citation
```bibtex
@misc{WeMM-Video,
title={WeMM-Video: Pushing the Frontiers of Open-Source Omni-Modal Video Understanding},
author={Changli Tang and Jie Yang and Tianyi Wang and Lei Li and Daoze Zhang and Zijie Xin and Zhan Liu and Li Li and Jinchang Xu and Fengyun Rao and Chao Zhang and Jing Lyu},
year={2026},
eprint={2611.26409},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2611.26409},
}
```