MarkTechPost publishes a Saluki 27B tool-calling test guideMachine translation
Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.
The guide runs the 7.9 GB Underdog Saluki 27B in llama.cpp and compares it with ISTA-DASLab’s quantized version on the same 120 tool-calling tasks. Its authors say Saluki’s weights also changed, so differences cannot be attributed to quantization alone; the test excludes the full-size model and is limited by its sample and single run.
This source permits summary display only.
Read at the original source