Underdog Releases Saluki 27B: A 2-bit Qwen3.8-27B GGUF That Beats the Original at Tool Calling

Conway Research has released Underdog Saluki 27B under Apache 2.0, a 2-bit GGUF of Qwen3.8-27B that fits in 7.89 GB versus 54 GB for BF16. It runs in stock llama.cpp with full GPU offload and outperforms the full model on tool calling: 88 versus 84 on Underdog Bench.

• 7.89 GB GGUF versus 54 GB for BF16, 27B dense parameters
• 96% average retention across 9 benchmarks
• Parallel tool calls: 42 versus 35 for the full model
• AIME 2025: 79.2 versus 96.7, about 82% retention