最新模型
Tencent Releases WeMM-Video for Joint Reasoning over Video and Original Audio; Deployment Still Requires Engine Patches
Tencent’s WeChat Vision team has released the WeMM-Video series, offering four model sizes and multiple quantized weights for joint audio-video understanding and Chinese and English responses. The technical report was submitted on November 19, but deployment requires specific inference environments; quantized file sizes also should not be treated as direct estimates of VRAM requirements.

Tencent’s WeChat Vision team has released WeMM-Video, which incorporates video frames and the original audio track into its understanding process. The technical report, submitted on November 19, covers 2B, 4B, and 9B dense models, as well as the 35B-A3B MoE, with approximately 35 billion total parameters and 3 billion activated parameters. The models accept text, images, audio, and video, and output text in Chinese and English. They can switch between direct-answer and step-by-step reasoning modes. Technical report
The technical approach uses visual and audio encoders to convert media into tokens, which are then processed by the language model; video frames carry timestamps. This lets an application combine visual events and dialogue in a single request, making the models suitable for testing audio-video summarization, content retrieval, and cross-modal question answering. The team says the models were pretrained on about 1.2 trillion multimodal tokens, followed by multi-teacher OPD, MrDPO to suppress repetition, and task-specific reinforcement learning. Model card
Developers can access BF16, FP8, AWQ, GGUF, and NVFP4 versions, but each format has a different execution path: Tencent provides patched versions of vLLM and SGLang, as well as a native runtime for GGUF. The service interface uses /v1/chat/completions; video_url.use_audio determines whether the original audio is read, while fps and max_frames control frame sampling. For engineering evaluations, sampling settings should be held constant to compare quality and cost with and without the audio track enabled. Deployment documentation
Memory planning requires particular care. The 35B-A3B model’s BF16 weights are about 67.39 GiB, and additional space must be reserved for media inputs and generation caches; 3 billion activated parameters should not be mistaken for the full model’s loading requirements. The 2B AWQ model card also states that its supplied Transformers loading method decompresses the quantized weights back to BF16, so a smaller download does not mean a proportional reduction in runtime VRAM use. Large model card, AWQ notes
The paper reports competitive results across more than 120 evaluations, but these are tests conducted by the release team. Before adoption, teams should retest dialogue recognition, audio-video alignment, and information loss in long videos using their own Chinese and English clips, while recording the reasoning mode, sampling rate, and latency. Independent reproduction of the results and upstream support in inference engines are the next developments to watch; both will affect deployment and maintenance costs. Evaluation summary