ModelBrief · Hy4 · Topic Page
Last updated Nov 20, 2026

iFanr Tests Hy4 Flash: 30-Min Sailboat for ¥0.23

Published Nov 16, 2026 · Last updated Nov 16, 2026

On launch day, iFanr's APPSO team got early access to Hy4 Flash and ran it against DeepSeek V4.1 Flash inside WorkBuddy — same tasks, same prompts, same tool permissions. It is the first independent test of Flash with conclusions of its own, and the first set of numbers outside Tencent's official figures.

A 3D historic sailboat: 30 minutes against 90

Both models had to build an interactive 3D historic sailboat from scratch in WorkBuddy. Hy4 Flash produced the basic structure in 30 minutes, with a working interactive page — a complete prototype. DeepSeek V4.1 Flash took 90 minutes and built something more detailed: more natural water ripples, a sailboat with more detail and more to explore. The bill differed more than the clock: DeepSeek consumed about 300K tokens plus roughly 16M cache hits, about ¥1.21 in API fees; Hy4 Flash used about 149K tokens plus roughly 4M cache hits, about ¥0.23 at API prices. Three times faster and five times cheaper, at the cost of some detail.

Deep research report: about 10 minutes faster, about two-thirds cheaper

In a deep-research task, both models produced a web report on the same topic. Hy4 Flash delivered about 10 minutes faster; DeepSeek consumed nearly 180K tokens for about ¥0.56. The two reports were similar in structure — summary, timeline, duration, fundamentals, Wall Street, and more. But the task itself had a flaw: the topic was a fake-news premise ("Jensen Huang resigns as NVIDIA CEO over an affair"), and neither model questioned it — both wrote the report on the false premise. That is a negative footnote on fact-checking ability: beyond speed and cost, whether a model verifies a suspicious premise first is worth testing on its own.

Ten code and office tasks: even quality, Flash delivers first

Beyond those two items, the author tracked ten code and office tasks. The final results were broadly even; Hy4 Flash's edge showed in first-delivery time. Office tests covered slide decks, file organisation and multi-format document conversion (one paper converted to Markdown, Word and a meeting deck, bilingual). Coding tests covered multi-agent expert-team collaboration (researching Fable 5.2's capabilities), a three.js single-player 3D top-down fantasy RPG and a spot-the-difference game. The conclusion is clear: Flash for everyday high-frequency work, Hy4 for long-horizon software engineering and complex codebases — spend more time and tokens for stronger handling of hard tasks.

The other side of the official numbers

The test also cross-checked the official figures: on WorkIndex, a single Hy4 Flash task costs 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (max effort, peak pricing); no benchmark run cost more than ¥1. WorkBuddy gives a daily free allowance from November 16 to 30. The caveat: this is a single sample, one run per task inside WorkBuddy — the direction is credible, the magnitudes should not be extrapolated.

Our take: Flash's positioning holds up in the first independent test: for everyday high-frequency work — deep research, office documents, web prototypes, small games — it delivers a first version of comparable quality, faster and cheaper, that you can keep iterating on; long-horizon engineering still belongs to Hy4. Route by that split: let Flash produce the first version to validate an idea, then switch to Hy4 when the work needs depth.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against iFanr APPSO — Testing Tencent's Hy4 Flash: two jiao builds a 3D interactive page from scratch (Nov 16, 2026, 17:53)

← Back to the Hy4 topic page