Heterogeneous prefill/decode isn't new: dev tried 70B prefill + 7B decode in 2023, 'entirely useless'

bingxu_ · x · 2026-09-03

Responding to the newly floated idea of using different models for prefill and decode, developer bingxu says they tried exactly this in 2023 — a 70B model for prefill and a 7B model for decode — and found it "entirely useless." The approach is not new, and prior practice suggests it works poorly.

Related event: Heterogeneous prefill-decode split sparks debate after engineer reveals 2023 trial(3 posts)→

Original post →

More from Models

Models channel →