qwen3.8-flash-next

Qwen3.8-Flash-Next (UD-Q3_K_XL, Qwen4 preview, 176.9B-A6B)

Qwen4-architecture preview, running on an unmerged llama.cpp PR. The only entry here that is not fully GPU-resident: experts in RAM, n-gram table paged off disk. Cold load is ~15 minutes and it thinks at xhigh by default, so expect it to be the slowest column on the grid by some distance.

show

Everything it built 4 result(s)

Bounce — the full game

desktopminimal-v1generatedtimed out10m 00s cleanup

2h 10mto build
40tool calls
75.9 KBwritten

play ▸open raw ↗run log pi transcript

Bounce — Level 1

desktopminimal-v1generatedtimed out5m 00s cleanup

1h 05mto build
28tool calls
29.3 KBwritten

play ▸open raw ↗run log pi transcript

Breakout

desktopminimal-v1generated1 js errortimed out

1h 00mto build
43tool calls
40.9 KBwritten

play ▸open raw ↗run log pi transcript

Snake, with one idea of your own

desktopminimal-v1generatedtimed out

1h 00mto build
44tool calls
30.7 KBwritten

play ▸open raw ↗run log pi transcript

pi parameters

context window	262144
max tokens	32768
reasoning	True
thinking	model default
thinking format	qwen-chat-template
extra pi flags	—

This model's block in models.yaml — the knobs tuned so the harness behaves for it.