A video model that listens
Per its technical report, WeMM-Video is built around a single design choice: its audio and visual tokens are interleaved in time order and stamped, so a sound and the event that produced it arrive in one sequence rather than in separate channels. The report describes an architecture built on video-SALMONN 2 and Qwen3.5 backbones, with a Whisper-based audio encoder. Whether that design holds up on real footage is exactly the kind of thing the published benchmarks cannot settle on their own, because every number attached to it so far comes from Tencent.
The family covers a useful spread of sizes. Each of the four sizes ships in quantized forms (FP8, AWQ, GGUF, NVFP4); the 4B NVFP4 checkpoint's weight files take 7.42 GiB, per its model card — a weight-file figure, not a total runtime-memory number. At the top, the flagship is 35 billion parameters in total with 3 billion active, and the official serving example spreads it across two GPUs through a WeMM-patched vLLM or SGLang. The "A3B" in the name describes how much of the model fires per token, not how much of it you have to store and load.
Read the claim at the right size
The launch language says "new open-source state of the art," and the supporting evidence is a technical report evaluating the models on more than 120 benchmarks. A benchmark count that large sounds thorough. It also means the specific comparisons doing the work in the headline — flagship ahead of two named proprietary models, the 9B ahead of a 30B-class open rival — are developer-reported results that no outside lab has yet reproduced; the desk found no independent evaluation. Tencent has done the one thing that makes verification possible: the evaluation and reproduction pipeline is publicly released with the family. Until someone outside Tencent runs it, the ranking stands as a claim with unusually good paperwork.
There is a second, quieter point in the license. The models are Apache 2.0, with none of the strings attached to some of this year's open releases — Tencent's own Hy4 license, for example, requires a separate license once an operator passes US$50 million in revenue or 100 million monthly users, and MiniMax H3's Community License limits where outputs may be displayed. For a team building video search, moderation, meeting analysis or archive tagging, that difference can matter more than a few benchmark points, because it decides whether the model can be used at all.
What it is not
WeMM-Video does not generate video. It watches, listens, describes, transcribes and reasons, and answers in text. Anyone filing it next to Hailuo, Wan or LTX as a creation tool is filing a video-understanding model among video generators. Its comparison set is the other omni-modal understanding models, and the closed APIs teams currently pay to summarize and search footage. Those workloads — search, summarization, moderation, transcription with the audio in context — are what a team can now evaluate on its own footage, under a license that permits commercial use without a revenue ceiling.
Sources
- [1] Tencent WeMM-Video model cards and collection, Hugging Face (repositories created November 17, 2026; license Apache 2.0; 4B-NVFP4 weight files 7.42 GiB; 35B card: mixture-of-experts, 35B total / 3B active; serving examples use tensor parallelism across two GPUs)Read source
- [2] WeMM-Video technical report, arXiv 2611.26409, "WeMM-Video: Pushing the Frontiers of Open-Source Omni-Modal Video Understanding," submitted November 19, 2026 (architecture, 1.2 trillion training tokens, 120+ benchmark evaluations, flagship and 9B comparisons — all team-reported)Read source
- [3] Tencent WeMM-Video code, evaluation pipeline and collection license (models licensed under Apache License 2.0), GitHubRead source
- [4] Official WeChat/Tencent announcement, November 20, 2026 (public release timing), as tracked by AICoder and KBlip release logs