Conway Research has released Saluki 27B, an on-device assistant model available under the Apache 2.0 license. This 2-bit GGUF version of the Qwen3.8-27B model is designed for local agents and requires 7.89 GB of storage, significantly less than the 54 GB needed for the full BF16 version.
The model was developed through a three-layer process. It utilizes the base Qwen3.8-27B architecture, incorporates quantization techniques from ISTA-DASLab, and includes a final optimization pass by Underdog to preserve tool-calling capabilities. The resulting file is compatible with stock llama.cpp and supports full GPU offload.
In performance benchmarks, Saluki 27B achieved a score of 88 in the Underdog Bench, compared to 84 for the full model. It also demonstrated higher efficiency in parallel tool calls, completing 42 tasks versus 35 for the original. Optional vision add-ons of 629 MB or 928 MB are available for the text-only base file.
Despite its performance in tool calling, the model shows limitations in specific areas. It retains approximately 82% to 85% of the original model's performance in competition math and multi-step reasoning tasks. Additionally, the model occasionally exhibits minor formatting issues in parallel-call responses.
