Tencent open-sources 9B video agent WeVisCut, rivals Opus 5

Tencent has released WeVisCut, a 9B parameter agent for turning narration scripts into multi-minute videos, in a paper on arXiv that reports comparable quality to Claude Opus 5 at nearly 33x fewer generated tokens per successful output minute.

The pitch is single-model economics for a workflow usually built from an orchestrator plus a stack of task-specific editors. The abstract describes a four-stage data pipeline that "combines editing inversion with verified replay to turn recovered editing intent into executable supervision," applied to 3,750 finished videos to yield over 130K training examples. The resulting policy, built on Qwen3.5 and operating in a stateful 12-tool workflow, scored 90.5 on the authors' own 100-task benchmark against 89.9 for Claude Opus 5.

The GitHub release is unusually complete for a video-editing paper: Apache 2.0 licensing, the 9B policy weights on Hugging Face, the trajectory-collection pipeline that produced the training data, and what the team calls "the Studio, a live timeline that stays editable after the cut." The authors argue the inversion step is necessary because finished videos "do not expose these decisions" directly.

The margins sit on a benchmark the same team designed, and the token-cost ratio reflects one scaffolding choice compared against another. Independent replication of either number has not landed yet.