Skip to content

v3.0.0 — Qwen3.8 on AMD Strix Halo: comparable evidence, not one misleading number

Latest

Choose a tag to compare

@hogeheer499-commits hogeheer499-commits released this 25 Aug 15:48
· 268 commits to 99186474cbd3eb81ec788901595948b3799e7b56 since this release

Qwen3.8 27B now has a dedicated Strix Halo decision layer that explains why 20.42 t/s measured here and public 22–65 t/s routes are not interchangeable claims.

What is new:

  • the guide-measured official Ollama route: 292.49 prompt t/s and 20.42 generation t/s over nine warm repeats;
  • image, tools, thinking, and exact retrieval through 50,059 prompt tokens on the measured stack;
  • separately labeled external GMKtec evidence reaching 261,130 evaluated tokens;
  • a route matrix for stock MTP, tuned ROCmFP4/DFlash, quality-first, and still-unverified performance leads;
  • a matched Qwen3.8 benchmark ladder and issue template so community attention can become comparable evidence;
  • a shorter README decision path while preserving the existing SEO/hook introduction;
  • public freshness validation, current traction evidence, responsible share copy, and an affiliate-ranking firewall;
  • SEO authority consolidation that makes strixhaloguide.com the readable project destination while GitHub remains the canonical technical evidence source.

Project web home: https://strixhaloguide.com/
Qwen3.8 decision and sources: https://github.com/hogeheer499-commits/strix-halo-guide/blob/main/QWEN38_STRIX_HALO.md
Qwen3.8 web answer: https://hogeheer499-commits.github.io/strix-halo-guide/qwen38-strix-halo/
Contribute a route: https://github.com/hogeheer499-commits/strix-halo-guide/issues/new?template=qwen38-route-report.md

Community and external results remain separate from first-party Beelink measurements. There are no affiliate links as of August 25, 2026; future monetized links must be labeled and commission does not determine ranking or conclusions.