Block a user
qwen4_exp B: feasibility spike on beast (GATE) — measure the decode rate before anything is ported
Epic: serve Qwen3.8-Flash-Next (qwen4_exp) — a 180B sparse MoE whose architecture is built for beast's shape
Gate result (#309): serve. C–H are unblocked.
Full evidence in #309; spec from #308 is on main as
doc/qwen4_exp-port-spec.md. Short version and what it changes.
The number
**86.7…
qwen4_exp B: feasibility spike on beast (GATE) — measure the decode rate before anything is ported
Recommendation: serve — and the placement plan in #307 is inverted
Measured on beast (2 × RTX 5090, 63.7 GB VRAM; Ryzen 9 9950X3D; 123 GB
DDR5), llama.cpp daef7b6 built for sm_120…