Tencent’s WeChat Vision team, in collaboration with Xidian University, has released WeVisCut, an autonomous agent designed to handle the entire workflow of long-form video production from narration to final cut. The system demonstrates that specialized post-training can allow a compact 9B parameter model to outperform massive frontier models like Claude Opus 5 in specific creative tasks while slashing computational costs by over 97%. This release highlights a growing trend in agentic AI: moving away from general-purpose brute force toward highly optimized, stateful tool execution pipelines.

Recovering Intent from Finished Projects

The core innovation of WeVisCut lies in its data engine, which utilizes a technique called 'Editing Inversion' combined with 'Verified Replay.' Instead of relying on synthetic data or manual annotations, the team recovered executable supervision from 3,750 real-world finished video projects. This process generated 134,771 training examples, allowing the model to learn complex editing decisions by observing successful outcomes and replaying them with production tools. The four-stage pipeline includes observation, bounded repair, and trajectory materialization, ensuring that the agent learns from actual production constraints rather than theoretical scenarios.

Stateful Execution Beats Open-Loop Workflows

WeVisCut operates on a shared production state architecture, distinguishing it from traditional open-loop workflows. The agent consists of three main components: an Orchestrator that selects legal tools, a Generator for script and asset creation, and a Critic that inspects artifacts against quality dimensions. When the Critic identifies failures, it routes targeted repairs that preserve successful work, feeding inspection evidence back into the policy for subsequent decisions. This closed-loop system allows the agent to 'think' and 'act' iteratively, enabling users to chat with the agent after the first cut to swap shots, change music, or restyle captions without re-shooting or re-processing the entire video.

Outperforming Frontier Models on Efficiency

In evaluations across 100 held-out tasks, WeVisCut achieved an overall score of 90.5, surpassing Claude Opus 5’s score of 89.9. More impressively, it accomplished this while using approximately 33x fewer generated tokens per successful output minute. The model demonstrated superior performance in Shot Composition Fidelity (SCF), Temporal Precision Rate (TPR), and Layout Harmony Consistency (LHC), though it slightly trailed Claude Opus 5 in Visual Narrative Smoothness (VNSM). Compared to native systems like Crayotter and VideoAgent, WeVisCut ran 4.1x to 5.0x faster at 6.4x to 10.8x lower cost, while also achieving higher human preference ratings for theme, narration, and visual impact.

Key Takeaways

  • WeVisCut is a 9B parameter agent trained on 134,771 examples recovered from 3,750 real video projects.
  • The system outperforms Claude Opus 5 on long-form video tasks with 97.0% fewer generated tokens.
  • Human evaluation preferred WeVisCut over native competitors on every dimension, with an overall MOS of 4.56.
  • The agent uses a closed-loop 'observe-think-act' cycle that allows for local repairs and interactive revisions.

The Bottom Line

WeVisCut proves that you don't need trillion-parameter giants to handle complex creative workflows; you just need the right data and a tight feedback loop. This is a massive win for cost-conscious developers who need high-quality video automation without the cloud bill shock.