Qwen3.8-27B MTP Grafted to Unsloth Saves RAM, Requires Thinking Mode

Nyghtbynger · reddit · 2026-08-23

The author grafted the MTP (Multi-Token Prediction) mode onto the Unsloth version of Qwen3.8-27B, as the official Unsloth releases do not ship with it. Tests on Vulkan indicate this grafting method saves RAM compared to using external files. However, the model performs poorly with Thinking disabled but recovers some intelligence on cultural questions when Thinking is enabled.

Original post →

More from Infra

Infra channel →