Slop TVNewsLatest
News

Tencent opens WeMM-Video, omni-modal models that hear and see

The WeChat Vision family runs from 2B to 35B-A3B, reads a clip's frames and soundtrack together, and ships with Apache 2.0 code and an evaluation pipeline.

Illustration: Tencent opens WeMM-Video, omni-modal models that hear and see
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Tencent's WeChat Vision team opened WeMM-Video, four Chinese-English models at 2B, 4B, 9B and 35B-A3B parameters, on GitHub and Hugging Face on November 19, 2026.
  • By Tencent's own evaluation, the flagship WeMM-Video-35B-A3B sets a new open-source state of the art on omni-modal video tests, beating proprietary Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 and narrowing the gap with Gemini-3.0-Flash.
  • The models read a clip's frames and its original soundtrack together, and Tencent says they are already deployed in WeChat services for live-stream commerce and short-form drama analysis.
  • The project code and the checkpoint model cards are Apache 2.0, with runtimes for Transformers, patched vLLM, patched SGLang and a native GGUF path.

Tencent has opened WeMM-Video, a family of omni-modal models that watch a video and listen to its soundtrack at once, releasing the weights, a Docker setup and an evaluation pipeline on GitHub and Hugging Face.

The WeChat Vision team published the technical report on November 19, 2026 (arXiv:2611.26409), co-authored with Tsinghua University, introducing a Chinese-English bilingual model family.

WeMM-Video takes text, images, audio and video in and writes text out. It is a video-understanding model rather than a video generator: it describes what happens in a clip, transcribes the speech, and reasons about what is seen and heard. For video inputs it can read the frames alone or the frames together with the original audio track.

The family comes in four sizes: 2B, 4B and 9B dense models plus a 35B-A3B mixture-of-experts. Each checkpoint ships in BF16, FP8 Block128, AWQ and NVFP4 weights alongside a GGUF build, served through Transformers, patched vLLM, patched SGLang or a native runtime. One request can carry up to four images, two videos and two audio items inside the context window.

Tencent trained the models on about 1.2 trillion multimodal tokens, with joint audio-visual data the largest slice at 45.4%, and reports testing them across more than 120 benchmarks.

In the report, Tencent says the flagship WeMM-Video-35B-A3B sets a new open-source state of the art on omni-modal video tests, beating proprietary models Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 and narrowing the gap with Gemini-3.0-Flash. On the report's own table, the 35B-A3B model in its slower thinking mode scores 86.4 on Video-MME with audio, against 83.6 for Seed2.0-Mini-0428 and 88.6 for Gemini-3.0-Flash. Tencent says it beats those two proprietary models on 16 of 17 general audio-visual benchmarks, and that the compact 9B model beats the larger open-source Qwen3-Omni-30B-A3B-Instruct. The comparisons are Tencent's own.

The dense checkpoints are built on Alibaba's Qwen3.5 backbones and the 35B-A3B on Qwen3.6-35B-A3B, so part of the family sits on a rival lab's open models. Audio runs through a Whisper-based encoder that compresses each 0.1-second window into one token, about 10 audio tokens per second of sound.

Tencent says WeMM-Video is already deployed across WeChat video services, including live-stream commerce, short-form drama analysis and interactive video.

Tencent's repository and the Hugging Face model cards both list Apache 2.0. Tencent flags patched vLLM and SGLang builds and trust_remote_code, so setup is heavier than a stock model, and the README's examples run at an 8,192-token context.

For creators, the draw is cost and audio. A self-hosted 9B or 35B-A3B model that reads a soundtrack can handle talking-head clips, lectures and live-stream replays that frame-only open models miss, if the reported gains hold on your own footage.

The weights, evaluation pipeline and Docker setup are live on GitHub and Hugging Face now; the 35B-A3B checkpoint is 67.39 GiB in BF16 and the 9B is 19.54 GiB.

Sources

  1. github.com - family sizes, modes, license, runtimes, Docker, evaluation pipeline
  2. arxiv.org - technical report, date, authors, training, benchmark claims
  3. arxiv.org - benchmark tables (Video-MME with audio scores)
  4. huggingface.co - model collection and checkpoint variants
  5. huggingface.co - model card, Apache 2.0 license, weight size
  6. huggingface.co - 35B-A3B weights size and license
  7. tokenpost.com - independent coverage, Nov 20, 2026
  8. explainx.ai - independent write-up and limits
  9. blockchain.news - independent coverage