The second independent hands-on, published the same day at 16:53 after the ifanr test. Zhidx ran Hy4 Flash through five task groups inside WorkBuddy: a Huang Renxun music video, a one-prompt "Sims 5", a 3D tennis match, a Beijing-landmarks animation, and a Blender pelican riding a bicycle. All five groups delivered — and every one of them left visible rough edges. The verdict is direct: the model's standout skill is chaining multiple steps and tools together; its clearest weaknesses are fourfold — fine detail, aesthetics, complex-structure reproduction, and self-checking.
The opener is a music video of Huang Renxun performing Jolin Tsai's "Sun Don't Set". Following an AI music-video tutorial, Zhidx had the model work through music analysis, art direction, character design, storyboarding and video production in sequence. Character design used the model's vision: a real photo of Huang Renxun was uploaded, and the model turned his hairstyle, glasses and black leather jacket into a chibi pixel character, plus an interactive preview page — five facing directions, twelve expressions, twelve actions — pulling out a green GPU chip when the lyric mentions "a postcard of love". The model then designed 50+ shots from the song's rhythm and lyrics and wrote animation code to assemble the video. Two problems show: fine detail is thin, movements are stiff and some cuts are not smooth; and the subtitles drift — despite an explicit instruction to build them to the lyric timeline and check sync on export, they never lined up with the music.
The second group tested complex front-end and 3D work. One prompt — "build a Sims 5 with visuals and gameplay" — produced a playable simplified version: an isometric house with three family members, live needs (hunger, energy, fun, social, hygiene), clickable furniture, autonomous living, pause, fast-forward and build modes, and a behaviour log. But it is far from the real Sims: the characters are simple cartoon figures, the furniture is basic geometry, and the room and visuals are crude. The 3D tennis match came out stronger: an interactive 3D game with court, stands and clubhouse, serving, rallies, scoring, surface types and camera switching all working — but with concrete bugs: a flagpole in an implausible spot, rackets that sometimes never touch the ball, and spectators fused with their seats.
The third group called a perspective-illusion skill from GitHub to make a 10-second Beijing-landmarks animation: the Temple of Heaven, the Forbidden City corner tower, the CCTV headquarters, China Zun and the Bird's Nest arranged by depth, the camera passing through each in turn, with a Chinese-flavoured soundtrack added automatically. Camera movement and scene transitions mostly follow the brief, but the buildings do not hold up: the Temple of Heaven and China Zun are recognisable, the CCTV "big underpants" is severely deformed, and the other landmarks are hard to identify. The same group also built a rainy-night convenience-store diorama under Mount Fuji — store, vending machine, bicycle, street lamp and guardrail, pink-purple neon, rain and wet-road reflections creating a Japanese-animation night mood, freely rotatable and zoomable — the better-looking of the two, though still short on craft and aesthetics next to strong AI 3D work. The last group drove Blender directly: via the Blender Python API, a roughly 16-second animation of a pelican riding a bicycle. The model built the pelican, the bicycle and the path, set up tracking and close-up shots as asked; writing the code took about 26 minutes and rendering another 20-plus, close to 50 minutes in total. The picture is rough: pelican and bicycle built from simple geometry, some frames nearly textureless, like unpainted clay models; the wings never touch the handlebars, though the webbed feet still pedal in turn.
The author sums the weaknesses as four: fine detail (pictures, models and materials mostly at the "recognisable" level), aesthetics (shots, colour and composition depend heavily on the prompt), complex-structure reproduction (buildings, bodies and machines deform as they get more complex), and self-checking (the model cannot spot its own problems). The long chain is the other side: from character design and storyboarding to animation and export, the model completes multi-step work and strings tools together. The piece also quotes two vendor-table figures not previously on this site: PPTGenBench 54.2 (preview 52.3) and PPTCoEditBench 74.0 (71.5); on harder software-engineering work the gap to the full model stays wide — Terminal-Bench 4.0 at 19.2 against 44.4, ProgramBench at 16.0 against 30.0. The blind-test figures (2.79 satisfaction, 6.7-minute median task time, ¥0.26 per task) already appear in our launch article and are not repeated here.
Reference sources: this article is an original rewrite produced by this site; facts and figures were checked against Zhidx — First hands-on with Tencent's Hy4 flash: a Huang Renxun MV, cheaper than DeepSeek (Nov 16, 2026)
← Back to the Tencent Hunyuan Hy4 topic page