Tencent assembles WeVisCut — an agent that edits entire videos autonomously

🎨 Tencent assembles WeVisCut — an agent that edits entire videos autonomously

Tencent's Qwen3.5-9B model handles the entire editing process: it plans, invokes tools, verifies results, and corrects only the failed parts. On a 100-task benchmark, WeVisCut scores 90.5, outperforming Claude Opus 5 (89.9) and Gemini 3.1 Pro (84.6), while using 33 times fewer tokens per minute of video and costing 10.8 times less than Crayotter.

🌍 Industry argument: A narrowly fine-tuned 9B model matches frontier performance with radically lower token and API costs. The editing inversion + verified replay pipeline converted 3,750 finished videos into over 130,000 examples.

👤 An example of an agent that closes the entire video production cycle from script to final video and fixes its own errors; the article and project are open. Caveat: The origin of the training videos is not disclosed — the corpus requires verification.

Source 1: https://tencent.github.io/WeVisCut/ Source 2: https://arxiv.org/abs/2610.38352