Hunyuan Hy4 Flash + WorkBuddy: faster, cheaper, and it writes Three.js
The short version
Hy4 Flash is the fast variant of Tencent's Hy4: 224B total parameters (including 64B Engram), 12B active, multimodal, 1M context, and noticeably cheaper per token than the full model. The big change is that it drops the old "deep thinking" phase — no long idle reasoning, which makes it much better for office work and high-frequency repetitive tasks.
Three hands-on tests
1. Physics demo (education)
Prompt:
Explore the relationship between mass and volume: use a balance and a graduated cylinder to measure different copper and iron blocks, then derive the ratio.
It finished in 8 minutes, covering weighing, cylinder readings, data logging and the conclusion — and the density ratio changed correctly when switching metals.

Same task: Hy4 Flash took 8 minutes and 6.01M tokens; the full Hy4 took 34 minutes; DeepSeek V4.1 Flash needed 52 minutes and 28.8M tokens. Faster and cheaper, for real.
2. Three.js 3D: jet engine
Prompt:
Build a complete, interactive Three.js visualization of an aircraft turbojet engine.
It rendered the blades and combustion chamber, with tail flame, RPM and temperature readings reacting on startup, plus orbit and zoom.
3. Creative replication: wood movable-type printing
Prompt:
Build a wood movable-type printing page: type letters, let the wood blocks fall and stack like bricks, then ink, press and flip the paper, turning a random layout into a savable poster.
The blocks flip, collide and pile up at the bottom, and pressing print pops out an ink-textured proof. Surprisingly complete for a flash model.

One bad case
On the keyboard replication task it got the shell right, but the number pad and arrow keys were misaligned — world knowledge is still the weak spot.
Verdict
- Good for: office work, table restoration, high-frequency repetitive tasks, simple 3D/interactive pages.
- Not for: detail replication that needs strong world knowledge.
- Price: free on WorkBuddy until Nov 30 — worth a trial.