Meet the Underdog Saluki 27B: A 2-bit Qwen3.8-27B That Beats the Original at Tool Calling
A 2-bit quantized Qwen beats the original on tool calling—but loses on math. What's the trade-off worth?
Why it matters
Saluki 27B demonstrates that aggressive quantization can specialize model behavior—excelling at tool use while degrading reasoning capability. For practitioners, this surfaces a real engineering decision: optimize for agent tool-calling reliability or preserve general reasoning.
The key facts
10 to knowSaluki 27B: 2-bit GGUF quantization of Qwen3.8-27B
Model size: 7.89 GB (vs. 54 GB original)
Beats original on tool calling benchmarks
Loses ground on competition math and reasoning tasks
Licensed under Apache 2.0
Published on MarkTechPost (community reporting, not vendor announcement)
Saluki 27B: 7.89 GB, 2-bit GGUF quantization of Qwen3.8-27B
Apache 2.0 license
Outperforms original on tool calling benchmarks
Original Qwen3.8-27B: 54 GB unquantized
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Underdog Saluki 27B is a 7.89 GB, 2-bit GGUF of Qwen3.8-27B under Apache 2.0. It beats the 54 GB original on tool calling but gives up ground on competition math and reasoning.