Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

WeMM-Video combines video frames and original audio for text answers

Tencent's WeChat Vision team documented a bilingual model family that reasons across video, audio, images and text, with model sizes from 2B to 35B-A3B. [1][2][3]

Model access
Technical report and public model repositories; separate checkpoint and dependency terms apply [1][2][3]

What the model does

The technical report describes dense 2B, 4B and 9B models and a 35B-A3B mixture-of-experts variant. Inputs can include video frames and their soundtrack together; outputs are text. An instruct mode produces direct answers while a thinking mode supports longer reasoning. The official repository links checkpoints and serving examples. [1][2][3]

Why it matters

Joint audio and frame processing can support questions whose answers depend on both what is visible and what is heard. Multiple sizes give researchers different resource and capability tradeoffs to evaluate. [1][2][3]

Who can use it

Researchers and developers building Chinese-English video analysis, media search or audio-visual question-answering systems. [1][2][3]

Key dates

  1. Tencent publishes the WeMM-Video technical report [1][2][3]

Ways to use it

Evaluate the relevant size on representative clips and check serving requirements before deployment. Review custom model code and pin a revision before enabling repository code execution. [1][2][3]

Model limitations

Performance comparisons are reported by the authors and were not independently reproduced here. A text answer can misinterpret video or sound. Model sizes are family designations rather than guarantees of total memory needs; these are understanding models, not video-generation products. [1][2][3]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source WeMM-Video model collection license and third-party notices
  2. Primary source Tencent WeMM-Video official repository
  3. Primary source WeMM-Video: Pushing the Frontiers of Open-Source Omni-Modal Video Understanding

Related profiles

Models & assistants

Macaron V1.1 Venti Pro on Alibaba Cloud

MindLab / Alibaba Cloud

The November 20 release ledger adds the Pro model separately from the Macaron variants listed on November 16. Alibaba documents OpenAI-compatible and Anthropic-compatible API access, streaming and adjustable reasoning effort. Customers must…

High confidenceEvidence checked

Models & assistants

Inkling 2 Mini

Thinking Machines Lab

The model reuses selected transformer layers during inference. Its official card lists Apache 2.0 licensing and local deployment support on specified hardware, alongside access through the Tinker platform.

High confidenceEvidence checked

Models & assistants

Ayb 0.5 Armenian model family

Metric AI

The release adapts the Qwen3.5 architecture through Armenian tokenizer expansion and language training, while retaining its vision components. The publisher links model repositories and an evaluation collection.

High confidenceEvidence checked