Qwen3-8B gets a KV-approximation add-on that halves prefill time without touching the model

teortaxesTex · x · 2026-09-11

A developer reproduced the KV-approximation scheme (described as "Encoder-Decoder" in DeepSeek-V4.1-Flash) on Qwen3-8B: by adding an approximate model on top without modifying the original weights, prefill time was cut roughly in half with comparable output quality.

The quote tweet further suggests normally trained decodes also work in the YOCO regime — meaning existing models can gain this speedup via an external approximation, no retraining required.

Original post →

More from Infra

Infra channel →