Underdog Saluki: Qwen3.8-27B in under 8GB with tool calling intact (huggingface.co)

The recently released Underdog Saluki model, Qwen3.8-27B, dramatically compresses the size of a powerful AI tool, fitting into a remarkable 7.89 GB while maintaining robust tool-calling capabilities. This model offers a competitive edge by achieving an 88 out of 120 task pass rate on the Underdog Bench, surpassing the full-size version (54 GB) which scored 84. It effectively retains 82 to 85% of the original model's computational strength, proving particularly effective at reasoning tasks—but shows minor weaknesses in letter-level instruction puzzles. This development is significant for the AI/ML community, as it demonstrates the feasibility of deploying high-performing models with reduced memory requirements, which can enhance accessibility and efficiency, especially for edge devices with limited resources. Furthermore, the Saluki model supports multimodal capabilities, including optional vision add-ons, allowing for image processing alongside text—all while operating within existing frameworks like llama.cpp. By streamlining model architecture while preserving performance, Saluki sets a new standard for practical AI applications, potentially opening doors for wider integration and use in diverse environments.