Is WeVisCut state of the art?Pending benchmarks.Announced Oct 28, 2026; no independent results yet.
Tencent WeChat Vision's 9B tool-using agent model for video editing, fine-tuned from Qwen3.5-9B with Xidian University co-authors. Given a knowledge article or narration script, it builds a multi-minute edited video one tool call at a time. Its 12 tools cover planning, narration, asset retrieval, templates, composition, music, inspection, repair and stitching. It is not a text-to-video generator: it needs Tencent's WeVisCut runtime, which runs the tool calls. Weights on Hugging Face; the GitHub license file says Apache-2.0.
The details so far
- Lab
- Tencent
- Announced
- Oct 28, 2026
- API
- Not yet · None stated; run locally
- Open weights
- Published · Hugging Face tencent/WeVisCut, not gated: a merged BF16 checkpoint plus a LoRA adapter
- Parameters
- About 9.41B (9,409,813,744 BF16 parameters in the merged checkpoint)
- Type
- Agentic video editing: a policy model that emits one tool call per step (about 12 per run) and a runtime that executes them; not a standalone text-to-video model
- Base model
- Qwen/Qwen3.5-9B, fine-tuned by Tencent (per the Hugging Face card and the GitHub LICENSE)
- Task
- A knowledge article or narration script in, a multi-minute edited video out
- Tools
- 12 editing tools: segment, plan, TTS, retrieval, template, compose, BGM, judge, reflect, probe and stitch, run by the WeVisCut Studio runtime with an editable timeline
- Training
- Editing inversion and verified replay on 3,750 finished videos yield 134,771 training examples; SFT and runtime-verified RFT, then Guided-OPD and RCF-DAPO
- License
- The GitHub LICENSE says Apache-2.0 for WeVisCut's code, parameters and weights, with Qwen3.5-9B (Apache-2.0) and other third-party components under their own terms. The Hugging Face repo has no LICENSE file; its card is tagged other and says only that the code is Apache-2.0, with weights and third-party components under their respective terms
- Code
- GitHub Tencent/WeVisCut: the Studio runtime, the 12-tool harness and the trajectory-construction pipeline
- Requirements
- Python 3.11 or later and an NVIDIA GPU with about 24 GB for the BF16 model; TTS, retrieval and rendering services are configured separately
- Authors
- Chenyu Mu, Wenming Weng, Qinrui Wang, Xingsong Ye, Hao Yu, Jun He, Chong Sun, Chen Li, Jing Lyu, Erkun Yang and Cheng Deng (Xidian University; WeChat Vision, Tencent)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 90.5 overall on Tencent's 100-task script-to-video benchmark (five runs), against 89.9 for Claude Opus 5, 84.6 for Gemini 3.1 Pro and 74.5 for its Qwen3.5-9B base, in the same tool environment.
Lab claim: about 33x fewer generated tokens and 73.3% fewer tool calls than Claude Opus 5 per successful output minute.
Lab claim: +6.8 SCF over Crayotter and +10.0 over VideoAgent, running 4.1x and 5.0x faster at 10.8x and 6.4x lower cost, with human raters preferring it (4.56 vs 3.48 overall).
Sources
PAPERWeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video (arXiv)arxiv.org/abs/2610.38352 · Oct 28, 2026LABWeVisCut project page (WeChat Vision, Tencent)tencent.github.io/WeVisCut/REPOTencent/WeVisCut on GitHubgithub.com/Tencent/WeVisCutREPOtencent/WeVisCut on Hugging Facehuggingface.co/tencent/WeVisCutREPOWeVisCut LICENSE (GitHub)github.com/Tencent/WeVisCut/blob/main/LICENSEPRESSMachine Delusions post on Xx.com/Machinedelusion/status/2116668287656333409 · Oct 31, 2026Tracking continues until an independent board lists it.