Qwen3.8-Flash-Next

Name: OpenWeights Terminal measured pairs
Creator: Dvidia
License: https://owterminal.com/cite

One quoted capture. Rank only lives inside a filter.

DGX Spark · EXL3 · EXL3

Decode

79.5

Prefill

—

tok/100W

33.1

A Cruz-style greedy 400-token code job runs 79.5 on my card. His 79 reproduced.

Native ExLlamaV3 on one Spark. 102.6 is draft-heavy repetitive code, not a blanket coding rate.

Officialreproduced