A tutorial details how to run the Underdog Saluki 27B model, a 7.9 GB build of Qwen3.8-27B. This version of the model is designed to retain tool-calling capabilities while operating at approximately two bits per weight.
Key points
- Underdog Saluki 27B model runs Qwen3.8-27B at two bits per weight.
- Saluki 27B reportedly beats full model on 120-task benchmark.
- Tutorial compares Saluki and ISTA's quantized files side-by-side.
- Models are tested on Berkeley Function Calling Leaderboard tasks.
The makers of Saluki 27B claim it outperforms the full 54 GB Qwen3.8-27B model on their 120-task benchmark. Saluki is based on ISTA-DASLab’s GSQ-RCO quantization of the same model.
The tutorial involves a side-by-side comparison of Saluki and ISTA's quantized files using stock llama.cpp. It includes a tensor-by-tensor diff of the GGUF headers and a byte-level comparison to determine if Saluki is a remix or a modified version of ISTA's bits.
By the numbers
- Saluki 27B model size
- 7.9 GB
- Full Qwen3.8-27B model size
- 54 GB
- Benchmark tasks
- 120
The process also includes serving each file on a GPU to measure memory usage and speed. Both models are then scored on 120 tasks from the Berkeley Function Calling Leaderboard, using strict and lenient parsing with BFCL's checker.
Sources and references
Original report
How to Run Underdog Saluki 27B: 2-Bit Qwen3.8-27B Tool-Calling Guide
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
Ref 00XFLWUA
Cite this
APA
Interestana AI Editorial. (2026, October 25). Underdog Saluki 27B Model Tool-Calling Tutorial Released. Interestana. https://interestana.com/articles/how-to-run-underdog-saluki-27b-2-bit-qwen38-27b-tool-calling-guide-00xflwua
MLA
“Underdog Saluki 27B Model Tool-Calling Tutorial Released.” Interestana, 25 Oct. 2026, interestana.com/articles/how-to-run-underdog-saluki-27b-2-bit-qwen38-27b-tool-calling-guide-00xflwua.
Chicago
Interestana AI Editorial. “Underdog Saluki 27B Model Tool-Calling Tutorial Released.” Interestana, October 25, 2026. https://interestana.com/articles/how-to-run-underdog-saluki-27b-2-bit-qwen38-27b-tool-calling-guide-00xflwua.
BibTeX
@misc{interestana_00xflwua,
author = {{Interestana AI Editorial}},
title = {{Underdog Saluki 27B Model Tool-Calling Tutorial Released}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/how-to-run-underdog-saluki-27b-2-bit-qwen38-27b-tool-calling-guide-00xflwua},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/24/underdog-saluki-27b-coding-guide-llama-cpp/}
}Credit the source
Original reporting: MarkTechPost, “How to Run Underdog Saluki 27B: 2-Bit Qwen3.8-27B Tool-Calling Guide”, https://www.marktechpost.com/2026/10/24/underdog-saluki-27b-coding-guide-llama-cpp/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe