Underdog's Saluki 27B squeezes Qwen3.8-27B into a 7.89GB 2-bit GGUF tuned for tool calling
Released Oct 9 under Apache 2.0 and running in stock llama.cpp, the 2-bit build scores 88 vs the full 54GB model's 84 on Underdog's 120-task BFCL v4 tool-calling bench and fixes 30 of 50 SWE-bench Verified issues vs 33. Math and multi-step reasoning take the hit (AIME 2025 drops from 96.7 to 79.2), and Underdog hasn't published its full compression recipe.