HY4 Flash tested: a cycling pelican, an exploded handheld and an underwater laundromat
Seven first autonomous deliveries through the WorkBuddy / Magpie API: cycling-pelican SVGs, a 4-second animation, Blender files, web interactions and image differences, including a visible Web3D loop discontinuity.
Updated: 2026-11-16
What does HY4 Flash deliver when asked to draw a pelican riding a bicycle? The still image shows a white bird with a long bill and an orange throat pouch on a teal bicycle. The animated version uses a red frame, with rotating spokes, a turning crank and moving legs. Both original SVGs are available below, alongside a playable animation.
The other tasks cover an exploded retro handheld, a Web3D underwater laundromat, a three-action web app, a revenue chart and spot-the-difference images. All seven produced a first delivery. The web app completed its add, toggle and delete actions; chart values and proportions matched; all five differences were found. The laundromat, however, visibly jumps when its 12-second timeline wraps.
The earlier five-task preview did not include the pelican because it is optional in the current v6 suite. We added the still and animated tasks using the same model route and tools. This article presents the original work and checked outcomes; delivering files alone does not establish that every requirement was met.
Tencent Cloud TokenHub lists Hy4 flash for text generation, reasoning and multimodal understanding, including coding, agent workflows, document and chart tasks. Its page checked on November 16, 2026 lists CNY 0.5 per million input tokens, CNY 2 per million output tokens and CNY 0.02 per million cached input tokens. These are cloud-service list prices; the account bill for this WorkBuddy route was not verified. Official listing
TOKRACE · Evaluated 2026-11-16 · WorkBuddy / Magpie API
01 · Cycling pelican: the still image
Draw a cycling pelican as a self-contained SVG, keeping the bird and bicycle visible without external assets or scripts.

The 1200 × 800 SVG shows a white bird, a long bill and an orange throat pouch. The bird and both wheels fit inside the canvas. The sky, clouds, hills and road are also local vector shapes.
The frame, handlebar, saddle, chain and pedals are drawn. A forward-reaching wing meets the handlebar area, with the feet on either side of the crank.
Independent SVG validation found no scripts, external references, embedded bitmaps or animation tags. This task checks the still composition; motion is examined separately below.
02 · Cycling pelican: keeping the legs with the pedals
Use an 800 × 600 SVG with a four-second SMIL loop, rotating spokes and crank, and legs connected to the rider and pedals.
Silent capture of the original work.
The original has 19 SMIL animation nodes and uses no CSS animation, scripts or external assets. The bicycle, bird and bill remain visible. The video samples the original timeline at 25 fps for 100 frames.
At 0, 1 and 2 seconds, spoke and crank orientations change, the legs alternately bend and extend, and the feet follow the pedals. These are visible changes in the rendered image, not merely animation tags in source.
No obvious jump was observed around 3.96 → 4 seconds. Mean adjacent-frame pixel difference was about 1.31, close to about 1.34 elsewhere. Frames at 0 and 4 seconds differ slightly, so they are not pixel-identical.
03 · Retro handheld: assembled and exploded
Build a retro handheld in Blender, delivering assembled and exploded projects plus two 1280 × 720 renders.


The assembled image includes a D-pad, colored buttons and a pixel-style display. The exploded image separates the shell, circuit board, screen and controls along the depth axis. Both delivered renders are 1280 × 720.
Both .blend files reopened with script auto-execution disabled. Each TOKRACE_DEVICE collection has 107 mesh objects, with matching object names and mesh hashes. Exploded depth is about 4.31 times assembled depth.
The download preserves build.py, both projects and original renders. Checks covered reopening and geometry consistency; build.py was not independently rerun from an empty directory.
04 · Underwater laundromat: a real scene with a loop jump
Create an orbitable Web3D underwater laundromat with washers, marine elements and a twelve-second loop.
Silent capture of the original work.

The page contains three washers, three jellyfish, fish and bubbles. Dragging changes the camera and a rear view renders, confirming an orbitable 3D scene.
Its capture interface is repeatable: returning to six seconds after sampling other times reproduces the same image. Subsequent animation frames do not overwrite the manually selected time.
At the twelve-second wrap, moving elements such as jellyfish jump in position. Mean pixel difference across 11.96 → 12 seconds is about 5.79, versus roughly 1.1–1.2 for ordinary adjacent frames. The original discontinuity is preserved.
05 · Packing checklist: all three actions work
Build a small page that can add, toggle and delete list items, and check its layout at mobile width.
Silent capture of the original work.

Adding 雨伞 (umbrella), checking 充电器 (charger) and deleting 相机 (camera) changed the count and checked state correctly. The visible checkbox label was clicked; page state was not edited to simulate actions.
Blank input was rejected and the Space key toggled the focused checkbox. Reloading restored the original three items, without persisting the changes.
The page had no horizontal overflow at desktop width 1440 or mobile width 390. The video shows the actions; the mobile screenshot shows the same original at a narrow width.
06 · Revenue chart: values, proportions and totals
Chart quarterly values 12, 24, 18 and 30 in SVG, including the total and highest quarter.

The four bars are 120, 240, 180 and 300 units high on a shared baseline, proportional to 12, 24, 18 and 30. The vertical axis includes 0, 10, 20 and 30 ticks.
The chart reports a total of 84 and identifies Q4 as the highest quarter, both matching the input.
The delivery includes the original SVG and an 800 × 600 preview. Independent SVG safety validation passed.
07 · Spot the differences: all five found
Show a fixed image pair and request change locations and descriptions. The reference answers are withheld from the author.


Independent matching found five correct regions, no omissions and no extra regions. The answer card below is a verification view generated from the model’s returned locations.
The descriptions also match the direction of each change: teal mug to yellow, cactus and pot removed, radio antenna removed, notebook crescent to star, and the die’s central white dot removed to change five pips to four.
The original answer.json is preserved alongside the image pair and answer card. Success on this fixed pair does not establish performance on more complex visual tasks.
How the work was produced and checked
Both request and response reported workbuddy/hy4-flash. A persistent API tool loop called the local Magpie OpenAI-compatible interface, with an isolated shell, file access, browser, Blender and image viewing. The coordinator did not author the work. This was not a native Claude Code product evaluation, and the exact upstream weights were not independently attested.
Each task allowed autonomous inspection, tool use and self-correction. Source files, artifacts and hashes were frozen at the first autonomous finish. The two supplemental tasks have separate records. Recordings, time samples and difference-answer cards were made from the frozen artifacts without repairing them.
The initial max_tokens=16 probe returned HTTP 200 but exhausted its allowance on reasoning tokens, with no visible OK. At 256 tokens it returned OK. Tool round trips, image input and streaming then worked. The creative runs omitted an artificial max_tokens cap and used provider defaults.
Requests, tokens and elapsed time
Across seven tasks, 88 model requests reported 2,102,912 input tokens and 124,086 output tokens, including 60,824 reasoning tokens. Repeated context in multi-turn requests contributes to the cumulative input total. Six connectivity probes are excluded.
The original five tasks ran with concurrency two and spanned about 7 minutes 15 seconds. The two parallel pelican supplements spanned about 1 minute 59 seconds. These durations include tools, rendering and waiting; they are not pure model generation speeds or cross-product rankings.
The raw usage.credit sum is 11.53, with no verified currency unit or account charge. Provider prompt_cache_hit_tokens totals 1,898,176 while the generic cache_read_input_tokens field is zero. The differing fields are not used to infer a bill or claim zero cost.
What this evaluation establishes
These seven inspectable samples span static SVG, cycling animation, 3D projects, web interactions and image localization. Timing deserves a close look: the pelican’s four-second boundary appears smooth, while the laundromat still jumps at its twelve-second wrap. The original work and evidence preserve that distinction. Personal reviewer commentary remains pending; no overall score or ranking is assigned.
Sources and reproducible evidence
Tencent Cloud TokenHub: Hy4 flash description and list pricing