Tencent Introduces WeMM-Video Models in Four Sizes
The Chinese-English family handles text, images, audio and video, with versions ranging from 2 billion to 35 billion parameters.

Tencent introduced WeMM-Video, a Chinese-English model family for understanding video alongside text, images and audio, with four versions and a public GitHub repository for developers and researchers.
The lineup includes 2B, 4B, 9B and 35B-A3B models. They support an instruction mode for direct answers and a thinking mode for step-by-step reasoning. Training used a corpus of about 1.2 trillion tokens, and the models were evaluated across more than 120 benchmarks.
The 35B-A3B model is described as achieving open-source state-of-the-art performance in omni-modal video understanding and reasoning. It also outperforms the proprietary Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 models, and narrows the performance gap with Gemini-3.0-Flash.
The 9B version outperforms the larger open-source Qwen3-Omni-30B-A3B-Instruct. The GitHub repository includes an evaluation and reproduction pipeline, guides and examples for English and Simplified Chinese.