SYLIN

Qwen3.8-27B Ridge

by Empero AI · Model · Language & code models

Empero published Ridge, an architecture-aware 3.7 bpw GGUF of Qwen3.8-27B: 11.7 GB versus about 50 GB full size

AccessOpen weightsNeeds~14 GBDownload11.7 GBSize27B parametersLicenseApache-2.0

Evidence

  1. Aug 18
    Keeps the sensitive Gated DeltaNet state values at high precision and compresses mid-stack feed-forward layers instead of flattening everything to 2 bits
  2. Aug 18
    Presenter says generic quants like IQ2, including Unsloth's, ignore the hybrid architecture, while Ridge holds the sensitive layers at high precision

Cost to run

Alternatives

Made from