Strix Halo benchmark: Swift 1.5 vs Unsloth quants of Qwen3.8-Flash-Next
deepu105 · reddit · 2026-10-26
The author benchmarks UkisAI's Swift 1.5 (claims 63% fewer thinking tokens, <1% accuracy loss) against Unsloth UD-Q4KXL and other Qwen3.8-Flash-Next quants on an Asus ROG Flow Z13 (128GB unified memory, 70W), across Halogen, gufo, and llama.cpp engines with MTP enabled. Since gufo couldn't load UkisAI's 4-bit GGUFs, the author requantized Swift Q80 to the same per-tensor types as Unsloth's version using UkisAI's imatrix (identical 111.3GB size).
Key results (single run, take with a grain of salt):
- Halogen native (v2 .hgn) scored highest: 15/17 in 160 min, decode 51-60 t/s
- gufo + Swift UD-Q4KXL requant: 13/17 but fastest overall at 136 min; Pi coding tasks took just 24 min/56k tokens
- llama.cpp lagged badly: prefill 300 t/s (vs 900-1400 elsewhere), similar scores at 2-3x the time
- gufo + Unsloth UD-Q4KXL hit 14/17 with best TTFT (20-27s)
Testing covered 11 Pi coding agent tasks plus math and LCB tasks, with speeds measured at 32k filled prompt.
More from Infra
- Burned $5,300 of AI tokens in 30 days on a $300 subscription plan — theteknosaur · 2026-10-26
- Jane Street, Hudson River Trading build data centers and back chip startups for AI — The Decoder · 2026-10-26
- Hamster wheel 'powers' local AI in a hilariously polished livestream-style setup — Sad_Measurement_7502 · 2026-10-26
- $10-12K build for local agentic coding: 1x 5090 vs 4x 3090 vs Mac Studio M5 Ultra — Shadmelor · 2026-10-26
- $1.5M local inference cluster running 240+ kWh/day is the new 'lottery winner' flex — TheZachMueller · 2026-10-26
- Gartner: 80% of neoclouds to cease independent ops by 2030, $4T at write-down risk — rohanpaul_ai · 2026-10-26