How to Run Underdog Saluki 27B: 2-Bit Qwen3.8-27B Tool-Calling Guide

Why it matters

This tutorial independently examines Underdog Saluki 27B, a heavily compressed build of Qwen3.8-27B whose makers claim it retains tool calling at roughly two bits per weight and slightly beats the full model on their own benchmark. Because Saluki is built on ISTA-DASLab's GSQ-RCO quantization of the same underlying model, the piece places the two GGUF files side by side in stock llama.cpp and asks a pointed question: is Saluki a genuine remix of ISTA-DASLab's bits, or a meaningfully changed model? The method is unusually rigorous for a hands-on walkthrough. It begins with a tensor-by-tensor diff of the two GGUF headers, then a byte-level comparison of the files themselves. Both builds are then served on a single GPU, with memory and speed measured, and scored on the same seeded Berkeley Function Calling Leaderboard tasks using BFCL's own checker under both strict and lenient parsing. Every download is pinned by size and hash, making the workflow reproducible. For readers, it offers a template for verifying bold quantization claims rather than taking vendor benchmarks at face value.