MTP Doesn't Speed Up MoE Models? Qwen3.8 27B Gets 2x Boost, 35B Unchanged
chibop1 · reddit · 2026-08-16
A Reddit user tests Lightning MTP on OMLX and finds it speeds up Qwen3.8 27B (dense) by 2x, but has no effect on Qwen3.6 35B (MoE). Detailed benchmarks are provided, questioning whether MTP is incompatible with MoE architecture.
More from coding & agent
- Kimi K3 generates its own experiment API for optimizer research — eliebakouch · 2026-08-16
- Inherent Labs unveils Faraday, a 27B AI Scientist for paper replication — ChenhaoTan · 2026-08-16
- Custom Dev Setups Often Disappoint; Traditional Tooling Stays Reliable — bigblueboo · 2026-08-16
- Redditor Proposes a 'Garbage Collection' System to Triage AI's Exploding Artifacts — dht · 2026-08-16
- DeepSeek Harness hits 100k GitHub stars in under 48 hours, outpacing OpenClaw — Hesamation · 2026-08-16
- Frontend Trend: AI Might Herald the Return of Pure HTML Websites — gethackteam · 2026-08-16