Input a knowledge article or narration script, and output a finished video several minutes long. The core is a policy model trained on Qwen3.5-9B: at each step, it reads the current structured editing state (chapters, shots, narration, assets, templates, music, rendering) and issues only one tool call—chapter segmentation, shot planning, narration synthesis, asset retrieval, animation template application, music selection, review, reflection, repair, and concatenation. Producing one finished video requires approximately 12 calls; the workflow is not hard-coded, and guardrails only reject illegal calls without making decisions for the model. Training data comes from "editing reverse-engineering + verification replay": 3,750 finished videos are reconstructed into executable editing decisions, yielding over 13 ten thousand training samples.
Likes 14 · Code Stars 13
Reasons
- Comprehensive availability: The 9B policy weights (hosted on Hugging Face as tencent/WeVisCut), the Studio runtime with the same 12 tools and state format used during training, and the data construction pipeline data_construct_pipeline that generates training trajectories are all open-sourced, with code under the Apache-2.0 license.
- The authors claim it significantly outperforms the base model in a unified tool environment, achieving performance comparable to Claude Opus 5, while using approximately 33 times fewer tokens per minute of generated video; they also built a benchmark with 100 tasks and 5 metrics to simultaneously measure editing quality and delivery reliability.
- Studio is continuously editable: It features eight synchronized tracks (chapters, shots, narration, subtitles, assets, templates, music, rendering), supports pause/resume, manual or chat-based edits (e.g., "change the opening," "replace the visual for 0:15"), re-renders only modified shots, and supports version rollback.
Usage
- Installation: git clone github.com/Tencent/WeVisCut → pip install -r requirements.txt (on GPU machines, first install PyTorch corresponding to your CUDA version, then pip install -r requirements-gpu.txt) → bash scripts/setup_external_tools.sh (fetches ffmpeg, Node, and the renderer, approximately 1.1 GB).
- Weights: hf download tencent/WeVisCut --local-dir checkpoints/weviscut-9b; bf16 inference requires an NVIDIA GPU with approximately 24 GB of VRAM, and Python ≥ 3.11.
- Launch: bash run_studio.sh (you can first use --check to verify paths), open http://<host>:8081/ in your browser, paste the narration script to start; use --no-model to view the interface without loading the model.
- Asset and music sources can be modified in studio_config.yaml to point to your own HTTP interfaces, local folders, or Python plugins.
Risks
- It is not a text-to-video model: Visuals come from asset retrieval and animation templates, and the quality of the finished video depends on the retrieval, TTS, and rendering services you configure; the model card states that factual accuracy and video quality are not guaranteed, and manual review is required before publication.
- Default assets/music are sourced from open web retrieval, and copyright responsibility lies with you; the Hugging Face weight license is marked as "other," and the README only states "weights and third-party components are subject to their respective licenses," so verification is required before commercial use (the base model is Qwen3.5-9B).
- Loading weights requires ms-swift with qwen3_5 registered; comparisons with Claude Opus 5 and token savings figures are reported by the authors, and the benchmark is self-built by the authors.
Editor's Note: Suitable for rough editing of knowledge-based and explanatory narration videos: let it generate a first version, then refine shot-by-shot on the timeline.