This month, Underdog AI, an artificial intelligence laboratory, has introduced the Saluki 27B model, a specialized variant tailored for agent-based applications. This model represents a 2-bit quantized and finely-tuned adaptation of the Qwen3.8-27B, streamlined to focus solely on text input and output functionalities, resulting in a remarkably compact file size of merely 7.89GB. While it demonstrates proficiency in everyday reasoning and tool utilization, its mathematical prowess is marginally less robust. Nevertheless, it surpasses the comprehensive Qwen3.8-27B model in terms of tool selection proficiency and the concurrent invocation of multiple tools. Saluki 27B adheres to the standard llama.cpp format, enabling seamless integration into standard llama.cpp environments and their associated applications, eliminating the necessity for custom compilation.
