The AI laboratory Underdog AI has introduced the Saluki 27B model. This new release is a 2-bit quantized version of the Qwen3.8-27B model, specifically optimized for text-based input and output tasks.
The model occupies 7.89 GB of storage space. Underdog AI notes that the architecture is tailored for agent-based applications, excelling in daily reasoning and tool utilization tasks, although it performs less effectively in mathematical computations.
According to the developer, the model demonstrates superior performance in tool selection and parallel tool execution compared to the full-sized Qwen3.8-27B. It utilizes the standard llama.cpp format.
Users can load the model directly into standard llama.cpp or any applications built upon this framework. No custom compilation is required to integrate the model into existing workflows.
