Use cases
codingagentic
Summary
User reports Qwen3.8 Flash-Next at 94.3 tok/s on one short coding request on a 128 GB Beelink Strix Halo.
Setup is Halogen 0.20.7 with HT4 weights; a separate run on 0.20.1 prefilled a fresh 32K prompt at 1,933 tok/s.
Combined throughput across four requests was 130 tok/s, and cached coding follow-up first-token latency fell from 620 ms to 131 ms. The user notes some figures beat published single-DGX Spark runs, with different prompts and quants.