Tencent's WeVisCut edits a long video from a script with a 9B model

Tencent Research has built an AI agent that takes a narration script and hands back a finished long-form video, running the whole edit through a single 9-billion-parameter model. The system, called WeVisCut, plans the shots, retrieves the footage, synthesizes the narration, composes the timeline, inspects its own cut and repairs only the segments that fail. On the team's own 100-task benchmark it scores 90.5 overall, narrowly ahead of Claude Opus 5 at 89.9, while spending about 33 times fewer generated tokens per finished minute. Tencent published the paper on October 28 and released the code and weights with it.

The question creators keep asking is whether a small, task-trained model can run a pipeline that usually needs a frontier model plus a stack of separate tools. WeVisCut is Tencent's answer, and because it shipped openly, anyone can check the claim.

What is WeVisCut?

WeVisCut is a long-form script-to-video system from WeChat Vision at Tencent and Xidian University. It works as an editor: give it a narration script, a target aspect ratio and a duration range, and it assembles a video from retrieved footage, text-to-speech narration, animated templates and background music, then stitches the shots into a final cut. It does not generate footage from a text prompt.

The whole thing runs on one policy model of 9 billion parameters, built on Qwen3.5-9B. The authors are Chenyu Mu, Wenming Weng, Qinrui Wang, Xingsong Ye, Hao Yu, Jun He, Chong Sun, Chen Li, Jing Lyu, Erkun Yang and Cheng Deng. The project page lists the two labs and links straight to the code and the weights.

How does WeVisCut edit?

The agent runs a stateful loop. At every step it reads the current editing state, which lists chapters, shots, narration audio, retrieved footage, chosen templates, background music and render status, and emits exactly one tool call. The released repository puts a rollout at roughly twelve tool calls. Twelve tools make up the whole action space: segment the article, plan shots, synthesize narration, retrieve visuals, fill a template, compose a shot, search for music, judge the text, judge the rendered video, reflect on what to repair, probe external services, and stitch the final cut.

The same weights serve three roles. An orchestrator picks the next tool, a generator writes script segments and shot plans and fills template content, and a critic inspects the returned artifacts and requests repairs. Guardrails reject illegal calls, but they do not choose actions for the agent.

The hard part was data. A finished video keeps the result and discards the decisions, the intermediate files and the failures, which is exactly what an agent needs to learn. Tencent's pipeline runs four stages: editing inversion recovers the script-to-shot intent from a finished video, verified replay executes that intent with production tools, an observe-reflect-repair step fixes defects without discarding work that already succeeded, and trajectory materialization turns the recordings into training examples. Applied to 3,750 finished narration videos collected from Bilibili, spanning technology, business, science and the humanities, the pipeline produced 134,771 examples.

Training then ran in three phases. The paper's own table shows the climb: the Qwen3.5-9B base scores 74.5 overall, and each stage lifts it, to 79.5, 83.5 and 84.9, before the final reinforcement step reaches 90.5. That is a 16.0-point gain over the base model.

What does the benchmark show?

Every number here comes from a 100-task benchmark Tencent built and five metrics the team defined. In a shared tool environment, WeVisCut scores 90.5 overall. Claude Opus 5 scores 89.9, Gemini 3.1 Pro 84.6, Qwen3.5-397B-A17B 81.5, Qwen3.6-35B-A3B 77.7, and Qwen3.6-27B 75.9. The base Qwen3.5-9B sits at 74.5.

The gap between WeVisCut and Claude Opus 5 is 0.6 points, and the paper's own standard deviations run one to two points, so the two land close to a tie. WeVisCut wins four of the five metrics and loses one, visual-narration semantic match, where it scores 74.6 against Opus 5's 74.8.

Against editors that run their own tools, WeVisCut leads: 90.5 overall against Crayotter at 84.9 and VideoAgent at 82.5. In a blinded human evaluation with three raters over the same 100 tasks, WeVisCut averaged 4.56 for overall preference against 3.48 for Crayotter and 3.46 for VideoAgent.

What does WeVisCut cost to run?

Cost is where the gap widens. Per successful output minute, WeVisCut generates 2,377 tokens; Claude Opus 5 generates 78,763, about 33 times as many. WeVisCut makes 3.1 tool calls per minute against Opus 5's 11.6.

In its native workflow, WeVisCut takes 2.1 minutes and $0.09 per output minute, against Crayotter's 8.6 minutes and $0.97 and VideoAgent's 10.4 minutes and $0.58. On the paper's own figures, that makes WeVisCut about 10.8 times cheaper than Crayotter ($0.97 divided by $0.09).

Every one of those numbers is the authors' measurement in the authors' environment.

What are the limits?

Read the results as claims, not settled fact. The benchmark is Tencent's own, and the paper does not disclose the scoring criteria or the exact conditions the baselines ran under. Independent replication has not landed, as AI Weekly noted when it covered the release.

The 0.6-point lead over Claude Opus 5 may sit inside the noise; with standard deviations of one to two points, a rerun could flip it. The paper is careful here, stating that the human-evaluation means do not establish statistical significance and do not validate the individual automatic metrics.

Other constraints are built in. Visual style is bounded by the HyperFrames template library, so uncommon looks get approximated with the closest available template. Verified replay guarantees the recovered edit runs only inside the registered runtime. One of the five metrics, the visual-narration match, is scored by a vision-language model, which can carry judge bias or drift as the service changes.

The paper's reference models, including the larger Qwen3.6-35B-A3B, received no task-specific training. Its 77.7 against WeVisCut's 90.5 compares a task-trained 9B against an untuned 35B, so it argues for specialization without proving that the same recipe would not do better at 35 billion parameters.

What it means for creators

For editors, WeVisCut points at a different shape for the work. The unit it reasons about is the shot, and it holds a production state across tool calls, so it can change one shot's timing without rebuilding the whole timeline. If the approach holds outside Tencent's benchmark, the round-trips that burn a frontier model's tokens, the re-planning, the re-rendering and the re-checking, could run on a model small enough to sit on one 24GB GPU.

That is a conditional. WeVisCut assembles videos from retrieved footage and filled templates; it does not shoot anything, and its look is limited by the templates it has. A creator who needs a specific visual style or original footage is not the target. The target is the assembly line: narration-driven explainers and knowledge videos, where most of the time goes into cutting and checking rather than shooting.

How to get WeVisCut

Tencent released the code on GitHub under Apache-2.0, the 9B weights on Hugging Face, and the Studio, an editable timeline that lets a person pause the agent, revise a shot by chat or by hand, re-render single shots and roll back to an earlier version. The repository also ships data_construct_pipeline, the trajectory-collection pipeline that built the training data.

The weights are 9 billion parameters in BF16, about 18GB, and inference wants an NVIDIA GPU with about 24GB of memory and Python 3.11 or later. The checkpoint is not useful on its own: it emits tool calls, and the runtime executes them, so it needs the WeVisCut environment to produce anything.

To try it, clone the Tencent repository, run scripts/setup_external_tools.sh, download the checkpoint, and start the Studio with run_studio.sh.