Underdog introduces saluki 27b: a 2-bit Qwen3.8-27B model that outperforms the original in tool calling

Underdog, a research firm based in Conway, has launched Saluki 27B, a 2-bit compressed version of Qwen3.8-27B, under the Apache 2.0 license. The model, which fits into just 7.89 GB, is designed for local agents and runs efficiently on standard llama.cpp. The compression technique was specifically tuned to preserve tool calling capabilities, a key feature for transforming chat models into functional agents. The full BF16 version of the model requires 54 GB of memory. Saluki 27B competes with models like Bonsai 2, which, despite being smaller, achieves 98.2% retention across 14 thinking-mode benchmarks and scores 95.00 on AIME25. However, Bonsai 2 requires a modified version of llama.cpp, unlike Saluki 27B, which works with the standard implementation. The release includes a Model Card on Hugging Face, a launch page, and an announcement on X.