The author reports that Qwen4-Max, in Agent goal mode, completed a recreation of a 23.8-second Three.js animation and then iterated through four more rounds of changes to produce a fifth version. The article records end-to-end task duration without separating model inference, tool calls, and rendering time.
Tasks this report can help assess: long-running coding and graphics tasks, multiple rounds of checklist-based changes, and workflows where an Agent captures screenshots and checks its own work.
Tasks this report cannot be generalized to: general coding success rates, task duration for other types of work, or rankings of model inference speed or cost.
Model version used: The author identifies it as Qwen4-Max but does not provide an API model ID, specific snapshot, or request logs.
Test environment or client: Agent goal mode; the client name and version, machine, and tool versions are not specified.
Reasoning level and parameters: The reasoning level and sampling parameters used in the test are not specified. The article discusses High and xHigh at the end but does not say which level this case used.
For the main case, no 3D model or project code was provided. The only input was a 23.8-second reference video, with instructions to recreate it in Three.js and render the finished video. The video was 1280×720, 30 fps, and 714 frames long. The author says the first round ran automatically without human intervention; afterward, the author reviewed the results and provided a checklist of changes for each round.
The article title says it "withstood 5 rounds of requested changes," but the body describes an initial build from scratch followed by four rounds of changes, from round two through round five, resulting in five versions. All durations below are task durations reported by the author:
| Stage | Duration reported by the author | Visible results and limitations |
|---|---|---|
| Round 1: build from scratch | More than 5 hours | Reconstructed camera angles and object motion from the video and produced the first version. The author says the camera position, composition, tonearm rhythm, and cat position were broadly aligned, but the details were rough. |
| Round 2: 12 requested changes | More than 5 hours | Added a curved tonearm, a starry central hole, small spring flowers, and ears and a tail for the cat. |
| Round 3: 6 requested changes | 2 hours 12 minutes | The author says 5 of the 6 were completed; lowered the autumn grass and revealed the hole, tonearm, and cat's path. |
| Round 4: 7 requested changes | About 3 hours | Adjusted the snow color, added a frost-forming animation at second 17, and changed the tonearm layers. |
| Round 5: 6 requested changes | About 2 hours | Adjusted the central structure and reduced how much the grass obscured the cat. The author says the final output matched the original video's specifications. |
A separate Blender case involved an 8-second rainy-night café animation at 1920×1080 and 24 fps (192 frames). The author reports completing it in 34 minutes. The input totaled 12.54 million tokens, of which 98.6% were cache hits. The article does not disclose the cost of this task.
The author reports analyzing all 714 frames in the main case and using geometric fitting, motion-path tables, screenshots, and pixel comparisons to assist generation and self-checking. The article mentions a camera-geometry error below 0.01 pixels but does not publish fitting data or measurement scripts for verification.
The Blender case's 12.54 million tokens and 98.6% cache-hit rate are input statistics, not model compute time. They also cannot stand in for token or cost data from the main case.
The article includes process images and video clips but does not provide the prompts from each round, the complete conversation or tool logs, a code repository, rendered files, or an independent evaluation.
This is a single author's case report, not a controlled benchmark. Durations such as "more than 5 hours" for the main case should be understood as end-to-end wall-clock time. They may include model waiting, Agent tool execution, script runs, and video rendering, and should not be described as pure model inference time. The article does not disclose the exact client and version, reasoning level used, machine configuration, service routing, exact inputs, or run logs. As a result, the task cannot be reproduced precisely, and the reported pixel error and final video quality cannot be independently verified. Round 3 also explicitly completed only 5 of its 6 requested changes.
A similar task could be recreated as described in the article: provide a reference video, ask the Agent to build a Three.js scene, render it, capture screenshots for verification, and then iterate through change checklists round by round. Exact replication would require the author to provide the original video, prompts from each round, project files, tool and client versions, model routing, and timestamped run records.
Qwen4-Max