Qwen3.8-27B-Splash-Mixed (Selective 8-bit Sensitive Layers)

This repository contains Qwen3.8-27B-Splash-Mixed, an experimental hybrid-precision model for the Splash inference engine on Apple Silicon.

Architecture

Based on mathematical sensitivity analysis (Hessian trace and singular-value decay), the top 8 deepest transformer layers (Layers 56–63), the token embedding table, and the LM head were selectively upgraded to uncompressed native 8-bit precision, while maintaining baseline compression on the earlier 56 layers.

Results

• Solves the 'Reasoning Cliff': Solved MATH-500 Problem 0 correctly (), matching full 8-bit Splash-HQ while saving significant memory.
• Decode Speed: 34.1 tok/s average (peaking at 48.6 tok/s), delivering a 3.44x speedup over standard autoregressive decoding (9.9 tok/s).