Qwen 3.8vsQwen 3.6vsHaiku 4.5

Each model received one prompt: build Flappy Bird in three files, with canvas only and no libraries. The code is not changed. Select a cabinet. Press Space or click to flap.

19m 12sQwen 3.8 build time 6m 35sQwen 3.6 build time 27sHaiku 4.5 build time 14,448 / 6,760 / 6,053tokens (3.8 / 3.6 / Haiku) M4 Maxllama.cpp, 36 GB

Qwen 3.8 27B alibaba · UD-Q4_K_XL

thought 13m 46s, wrote 5m 26s 12.5 tok/s3 files

Qwen 3.6 27B alibaba · UD-Q4_K_XL

typing from 21s in, done in 6m 35s 17.1 tok/s3 files

Haiku 4.5 anthropic · claude api

done in 27s flat 6,053 tokens3 files

Each model made its game in one attempt, from the same prompt. The two Qwen models ran fully offline on a MacBook Pro, with llama-server and the Q4_K_XL quant. Haiku 4.5 ran in the cloud through the Claude API, at a cost of $0.08. We applied one patch: an in-memory localStorage shim, which the page sandbox makes necessary. Only the game from Qwen 3.6 uses localStorage.