WeMM-Video-35B-A3B · AWQ W4A16
Overview
This repository contains the AWQ W4A16 Safetensors variant of WeMM-Video-35B-A3B. WeMM-Video is a Chinese–English bilingual family for joint audio-visual understanding and reasoning. It accepts text, images, audio, and video and generates text responses.
- Weights · 22.69 GiB of weight files
- Runtime · patched vLLM 0.18.0 with the checkpoint's compressed-tensors configuration
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 for family details.
Quantization
Routed and shared expert feed-forward weights use group-128 INT4 AWQ in compressed-tensors format. Activations and attention weights use BF16. See the quantization manifest for tensor scope and export checks.
Direct loading with Transformers
Use the Transformers environment. The local release checkpoint passed full weight loading and generation on 2 NVIDIA L20 GPUs with Transformers 5.2.0 and PyTorch 2.10.0+cu129. Generation checks covered text, image, audio, video frames, and video with its soundtrack. The same loader accepts the model's Hub ID. The validation used local checkpoint files.
Install compressed-tensors==0.13.0. The in-memory target override selects the packed feed-forward tensors. The loader decompresses them to BF16. Budget memory for the decompressed weights.
The tested device map assigns decoder layers 0–18, token embeddings, rotary embeddings, and media encoders to CUDA device 0. Layers 19–39, final normalization, and the output head run on CUDA device 1. The explicit map keeps Accelerate’s input hooks aligned with the assigned weights. Peak PyTorch allocations during these short generation checks were 41.13 and 42.19 GiB. The test harness accelerated full-name module lookup during loading with equivalent matches. It restored the original matcher before generation.
import torch
from transformers import AutoConfig, AutoModelForImageTextToText, AutoProcessor, CompressedTensorsConfig
model_id = "tencent/WeMM-Video-35BA3B-AWQ"
config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
targets = [r"re:^model\.language_model\.layers\.\d+\.mlp\.(?:(?:experts\.\d+|shared_expert)\.)?(?:gate|up|down)_proj$"]
for group in config.quantization_config["config_groups"].values():
group["targets"] = targets
device_map = {
"model.audio": 0,
"model.visual": 0,
"model.language_model.embed_tokens": 0,
"model.language_model.rotary_emb": 0,
**{
f"model.language_model.layers.{i}": int(i >= 19)
for i in range(40)
},
"model.language_model.norm": 1,
"lm_head": 1,
}
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
model_id, trust_remote_code=True, dtype=torch.bfloat16,
config=config,
quantization_config=CompressedTensorsConfig(run_compressed=False),
device_map=device_map,
).eval()
messages = [{"role": "user", "content": [
{"type": "text", "text": "What is 2 + 2? Reply with the number only."},
]}]
inputs = processor.apply_chat_template(
messages, tokenize=True, return_dict=True, return_tensors="pt",
add_generation_prompt=True, enable_thinking=False,
).to(device=model.get_input_embeddings().weight.device, dtype=torch.bfloat16)
with torch.inference_mode():
output = model.generate(**inputs, do_sample=False, max_new_tokens=32)
print(processor.tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))
Transformers input types
For a media request, place the media item before the text item in messages[0]["content"] and run the same preprocessing and generation.
| Input | Example content item |
|---|---|
| Text | {"type": "text", "text": "Describe what you see."} |
| Image | {"type": "image", "image": "image.jpg"} |
| Audio | {"type": "audio", "audio": "speech.wav"} |
| Video frames | {"type": "video", "video": "clip.mp4"} |
| Video with soundtrack | {"type": "video_audio", "video": "clip.mp4"} |
See the Transformers examples for media decoding dependencies and sampling options.
Serving
Install the matching WeMM runtime patch. The serving guide lists the dependency profiles and Docker images. Choose one server below.
vLLM
Use patched vLLM 0.18.0 with the checkpoint's compressed-tensors configuration. See the vLLM recipe for setup details.
MODEL=tencent/WeMM-Video-35BA3B-AWQ
vllm serve "$MODEL" --served-model-name "$MODEL" --trust-remote-code --dtype bfloat16 \
--host 127.0.0.1 --port 8000 --max-model-len 8192 \
--no-enable-prefix-caching --mamba-ssm-cache-dtype float32 \
--limit-mm-per-prompt '{"image":4,"video":2,"audio":2}' \
--media-io-kwargs '{"video":{"fps":2,"min_frames":4,"max_frames":16,"total_pixels":262144}}'
Input examples for the serving API
With the server running on port 8000, use the bundled request client from the project root in a second terminal. It sends media inline to the OpenAI-compatible chat endpoint.
MODEL=tencent/WeMM-Video-35BA3B-AWQ
python examples/chat.py text --model "$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.
Chat Modes
The checkpoint template defaults to Instruct. Set enable_thinking=True through the runtime's chat-template options to use Thinking mode. Use the media processor and chat template shipped with this checkpoint.
Validation
Patched vLLM loaded the checkpoint on H20 and completed generation checks for text, image, audio, and video with a soundtrack. The runtime precision audit confirmed INT4 feed-forward weights and BF16 attention.
Evaluation
See the WeMM-Video-35B-A3B evaluation results for the benchmark comparison in the technical report.
License
Apache License 2.0. See the included license and the collection license and third-party attribution. Third-party components remain subject to their original terms.
Citation
@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},
}
- Downloads last month
- -
Model tree for tencent/WeMM-Video-35BA3B-AWQ
Base model
tencent/WeMM-Video-35BA3B