WeVisCut: Tencent Openly Releases a 9-Billion-Parameter Video Editing Agent
Researchers from WeChat (Tencent) and Xidian University have released WeVisCut, a 9-billion-parameter agent that turns text or a narration script into an edited multi-minute video, with the code and editor licensed under Apache 2.0 and the weights hosted on Hugging Face. On their own benchmark, it matches Claude Opus 5 while generating roughly 33 times fewer tokens.
Tencent is open-sourcing a complete video editing agent. Researchers from WeChat Vision and Xidian University have released WeVisCut, a 9-billion-parameter model that transforms an article or narration script into a fully edited, multi-minute video. The code, editor, and data generation pipeline are released under an Apache 2.0 license, and the weights are hosted on Hugging Face. Their exact licensing remains to be clarified: the GitHub repository’s license file covers them under Apache 2.0, but the Hugging Face model card refers to separate terms for the weights and third-party components.
A Single Small Model Instead of an Agent Swarm
Most AI-based “editors” chain together multiple agents steered by large frontier models. WeVisCut takes the opposite approach: a single model acts in turn as orchestrator, writer, and critic. At each step, it assesses the project state and calls one of twelve tools: segmenting text into chapters, planning shots, generating voiceover audio, searching for images, populating animation templates, picking background music, rendering a shot, reviewing the output, self-correcting, and then assembling the final cut. This typically requires about a dozen calls per video.
The output opens in a dedicated editor called Studio, featuring eight synchronized tracks (chapters, shots, narration, subtitles, images, templates, music, rendering). Users can tweak a shot manually or via conversational prompts (“replace the image at 0:15”) and re-render only the affected portions.
Training from Finished Videos
The key challenge, the authors explain in their research paper, is that a finished video leaves no trace of the underlying editing decisions. To address this, the team “disassembled” 3,750 narrated videos published on Bilibili (covering science, technology, finance, history, and more) to reconstruct the intent behind every shot, then replayed that editing process using their own toolset—mistakes and corrections included. This yielded 134,771 training examples, which were used to fine-tune Qwen3.5-9B across three phases, including two reinforcement learning stages.
On Par with Claude Opus 5 Using 33 Times Fewer Tokens
According to the benchmark tables in the paper, on a suite of 100 tasks created by the team, WeVisCut scores an overall 90.5, compared to 89.9 for Claude Opus 5 using the same tools. Crucially, it generates 2,377 tokens per minute of successful video, compared to 78,763 for Opus 5—roughly 33 times fewer. When stacked up against two other research systems, Crayotter and VideoAgent, it also leads across all five metrics and in blind evaluations by three human annotators. Running the model requires around 24 GB of GPU memory on NVIDIA hardware.
Our take
WeVisCut targets a very specific format: the voiceover-driven explainer video illustrated with stock visuals, a style ubiquitous across YouTube and Bilibili. In that arena, it demonstrates that a compact, specialized model can rival a frontier model at a fraction of the inference cost while fitting onto a single GPU. Two caveats stand out: the benchmark and metrics were designed by the authors themselves, and the final visual quality hinges on open-web search results, raising copyright considerations for commercial applications. For creators, the primary immediate appeal is the editor, which keeps projects fully adjustable after the initial automated cut rather than locking them into a static video file.
Sources
- arXiv — WeVisCut: The Whole Editing Room in One Agent, From Narration to Long-Form Video (2610.38352) Primary source
- GitHub — Tencent/WeVisCut Primary source
- Hugging Face — tencent/WeVisCut Primary source
- GitHub — Tencent/WeVisCut, LICENSE file Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
