The author says he used the same prompt for Hy4 GA and an August 31 run of Hy4 Preview. He found the GA version's pedaling motion and foot placement more convincing, but flagged problems with the beak and sunglasses. The post lists a 19-second video. We could verify neither the footage nor the full prompt.
Tasks this can inform: The author's report of one same-prompt pelican animation comparison and his subjective observations about motion naturalness and geometric representation.
Tasks this cannot be generalized to: Other animation, video, or spatial generation tasks; the overall capabilities of either version; or general comparisons with DeepSeek.
Model versions covered: Hy4 GA; Hy4 Preview (the author says it was tested on 2026-08-31); DeepSeek V4.1 Flash 1017 (only a subjective point of reference from the author; this note contains no same-prompt comparison result).
Test environment or client: Not specified.
Reasoning tier and parameters: Not specified. Apart from the claim that the prompt was “exactly the same,” neither the prompt text nor generation settings were disclosed.
The author says he compared Hy4 GA with an August 31 run of Hy4 Preview using exactly the same prompt for a “pelican test.” The post describes spatial and geometric structure, world understanding, and the pedaling motion. It gives no prompt text, number of runs, client, or evaluation procedure. We did not verify any video frames.
Spatial and geometric structure: The author considers the GA version a clear improvement over Preview.
World understanding: The author also considers this a clear improvement, but gives no specific criteria or examples.
Motion and appearance: In a reply, the author said Preview looked better but missed the foot placement. GA aligned the feet and legs more accurately, while the sunglasses and beak still looked wrong.
Cross-model judgment: The author subjectively felt that this result was inferior to DeepSeek V4.1 Flash 1017. The original post said he would later compare DeepSeek and Kimi K3.1 Pro, so there are no verifiable same-prompt cross-model results here.
Media check: Tabbit showed a 19-second video on the original post. We did not verify the footage. The benchmark chart is from a quoted post.
| Item | Hy4 GA | Hy4 Preview |
|---|---|---|
| Test content | Pelican test, according to the author | Same-prompt test, according to the author |
| Prompt | The author says it was exactly the same as for Preview; text not disclosed | The author says it was exactly the same as for GA; text not disclosed |
| Test date | Not specified | 2026-08-31, according to the author |
| Spatial, geometric, and world understanding | The author says there was clear improvement | Comparison baseline; no separate description provided |
| Pedaling motion | Author reports more natural movement and more accurate foot and leg placement | Author reports inaccurate foot placement |
| Appearance | Author reports problems with the sunglasses and beak | Author preferred the modeling and appearance |
| Run settings | Not specified | Not specified |
The two runs span August 31 to November 4. The author gives no indication that they used the same client, parameters, or process. The visual differences are self-reported. The reference to DeepSeek V4.1 Flash 1017 is the author's impression, and the post gives no same-prompt result for that model.
Open the original post and replies in Tabbit to verify the author's assessment. The post showed a 19-second video label at collection, and we did not verify specific frames. To reproduce the comparison, obtain the original prompt and complete generation records for both versions, then hold the client, parameters, and process constant across new runs.
Hy4