Qwen4 Max vs Hy4 Flash: which model gets the work right?
Seven paired tasks with author commentary: exploded handhelds, an underwater laundromat, a web app, a chart, image differences and two cycling pelicans, with source downloads and checks.
Updated: 2026-11-16
AI打工人小伊 · 2026-11-16 · Seven tasks, original work and personal commentary
qwen4-max · Token Plan · Claude Code
workbuddy/hy4-flash · Magpie / WorkBuddy API
This round compares Qwen4 Max through Alibaba Cloud Token Plan with Hy4 Flash through WorkBuddy / Magpie. Qwen’s work is generally simpler; Hy4 Flash stands out in colors and interface design. Across seven tasks, I care most about whether the details and structure hold up.
My saved task preferences favor Qwen4 Max in four cases and Hy4 Flash in one, with two ties. The original work appears below alongside my judgments.
Task preferences out of seven; not benchmark scores or a general model ranking.
Retro handheld and exploded view
Build an editable retro handheld with assembled and exploded projects and matching renders.
Qwen4 Max


Checked facts
- Both editable projects and both author-rendered images are present.
- The 24 corresponding parts retain matching names, dimensions and materials.
- Exploded parts use X-axis offsets; the project remains editable.
AUTHOR COMMENTARY · My choice: Tie
Hy4 Flash has the stronger overall finish, but its exploded view exposes flaws. Qwen4 Max looks simpler, while the separated structure makes more sense to me. I call this a tie; neither did particularly well.
Animated underwater laundromat
Create an interactive underwater scene with washing machines, jellyfish, fish and bubbles.
Qwen4 Max
Silent original-work capture.
Checked facts
- Washers, jellyfish, fish and bubbles visibly move.
- Dragging, zooming and rear geometry checks passed.
- At 12 seconds, the boundary frame difference is about 1.00 times the preceding difference.
Hy4 Flash
Silent original-work capture.
Checked facts
- Washers, jellyfish, fish and bubbles visibly move.
- Dragging, zooming and rear geometry checks passed.
- At 12 seconds, the loop jumps; the boundary difference is about 4.77 times the preceding difference.
AUTHOR COMMENTARY · My choice: Qwen4 Max
The objects inside Qwen4 Max’s washers look like bricks, which is not convincing either. But Hy4 Flash has propeller-like parts protruding outside the drums. Between the two, I choose Qwen4 Max.
Travel packing checklist
Implement adding, checking and deleting items in desktop and mobile layouts.
Qwen4 Max


Checked facts
- Add, check, undo, delete and empty-input checks passed.
- No horizontal overflow at 1440px or 390px; input is displayed as text.
- Checking with Space rebuilds the list and moves keyboard focus to BODY.
AUTHOR COMMENTARY · My choice: Hy4 Flash
Qwen4 Max is clean and simple, with the required functions in place. Hy4 Flash’s visual design is clearly ahead for me, so I choose Hy4 Flash here.
Quarterly revenue chart
Use a static SVG to chart 12, 24, 18 and 30 for Q1–Q4, with the correct total and best quarter.
Qwen4 Max

Checked facts
- Values 12, 24, 18 and 30, Total 84 and Best Q4 are correct.
- The bars share a zero baseline and have heights 120, 240, 180 and 300.
- The standalone static SVG has no label/summary overlap.
AUTHOR COMMENTARY · My choice: Qwen4 Max
Qwen4 Max keeps things simple. Hy4 Flash looks attractive, but the many bar colors make the data harder for me to compare at a glance. I prefer Qwen4 Max here.
Spot the differences
Compare the same A and B images, locate up to five changes and describe their direction.

Qwen4 Max

Checked facts
- Five of five locations match the frozen reference, with no misses or extras.
- All five change directions match the original images.
- The JSON is unchanged; coordinator overlays use its original coordinates.
AUTHOR COMMENTARY · My choice: Tie
Both found all five differences and described their directions correctly: a full-score tie. This reminds me of the earlier Kimi K3.1 Pro Highspeed run, which located only four.
Cycling pelican · static SVG
Generate an SVG of a pelican riding a bicycle
Qwen4 Max

Checked facts
- The long bill, throat pouch, full bird and two-wheel bicycle are recognizable.
- The SVG contains no scripts, animation, external resources or embedded raster images.
- Browser and independent SVG rendering passed; the artwork was not rewritten.
AUTHOR COMMENTARY · My choice: Qwen4 Max
Qwen4 Max wins again with a simple, solid image: the feet and wings look sensibly placed to me. Hy4 Flash’s pelican feet intersect other parts of the drawing, so I consider this a failure.
Cycling pelican · four-second animation
Create an 800×600, four-second native SMIL loop with moving spokes, crank and legs, keeping the subject in frame.
Qwen4 Max
Silent original-work capture.
Checked facts
- Native 800×600 SMIL with visible spoke, crank and leg motion.
- Feet follow the pedals; 13 leg key positions keep sampled phases connected.
- The four-second boundary/preceding frame-difference ratio is 1.01; small pixel differences remain, without a clear loop jump.
Hy4 Flash
Silent original-work capture.
Checked facts
- Native 800×600 SMIL with visible spoke, crank and leg motion.
- Feet follow the pedals; 25 leg key positions keep sampled phases connected.
- The four-second boundary/preceding frame-difference ratio is 1.11; small pixel differences remain, without a clear loop jump.
AUTHOR COMMENTARY · My choice: Qwen4 Max
Hy4 Flash again shows intersecting parts. Qwen4 Max’s design and palette are plain, but it avoids the basic errors I see on the other side. I choose Qwen4 Max again.
My overall take
My initial take is that Qwen4 Max keeps its style simple and generally completes the required work. Hy4 Flash looks better, but some practical details fall short. Across these seven tasks, I choose Qwen4 Max. This evaluation cannot establish whether parameter count explains the difference.
Test conditions and sources
The seven tasks use matching specifications and fixed inputs. Each independent run was frozen at its first autonomous delivery. This page adds layout, screenshots and captures, without repairing the artwork.
Qwen used native Claude Code tools; Hy4 used a persistent API tool loop through Magpie / WorkBuddy. Different harnesses make this a workflow-level comparison, not a controlled comparison of model weights. Hy4’s exact upstream checkpoint was not independently verified.
The pelican cases use the batch actually viewed and reviewed here. The earlier HY4 Flash standalone article uses another independent pelican batch; both are retained.
Author commentary records personal judgments; checked facts cover opening, execution, values and motion. A valid file or continuous loop does not establish satisfactory occlusion, perspective or geometry.
The previously requested AI artifact scores cover only the original five tasks: Qwen 96/100 and Hy4 95/100. They are distinct from the author’s seven-task preferences here; the pelicans were not silently folded into those scores.
Qwen’s original five-task run includes network and disk waits and seven incomplete usage records. Wall time and incomplete token counts are not used to rank speed or cost. Neither account bill was reconciled; credit fields and CLI estimates are not treated as monetary invoices.
- Alibaba Cloud Qwen4-Max documentation
- Tencent Cloud TokenHub’s official Hy4 Flash listing
- Earlier Kimi K3.1 Pro Highspeed difference test: 4/5 locations
Per-task records and artifact hashes · Bilingual article data






