AnalysisAI ModelsOctober 26, 2026

Qwen4 leaks point to 27B model with 4-way residual stream

Read original source →reddit.com

A vLLM-XPU-kernels PR (since 404'd) described a 4-way parallel residual stream in Qwen4-27B to boost information flow within a fixed 27B parameter budget. A separate T-head PR referenced "Qwen4.0-Max prefill heads"; Alibaba has confirmed Qwen4 is in training with no release date.

2 sources

More stories today

Open the live feed