AI News Feed
Market watch
Cognition Open-Sources Devin Vulnerability Fixes Plugin for PHP, Node.js and GitLab RCE FlawsTutorial Tests Underdog Saluki 27B's 2-Bit Tool-Calling Claim Against ISTA QuantizationNew Zealand faces Cook Islands in crucial Rugby League World Cup matchSwitching Fitness Trackers in 2026 Can Preserve Old Health Data, but Cross-Brand Moves Need Manual ExportsTving Tops September Streaming App Installations Despite Data Breach ControversySamsung, SK hynix to report Q3 earnings with AI memory demand in focus
Cognition Open-Sources Devin Vulnerability Fixes Plugin for PHP, Node.js and GitLab RCE FlawsTutorial Tests Underdog Saluki 27B's 2-Bit Tool-Calling Claim Against ISTA QuantizationNew Zealand faces Cook Islands in crucial Rugby League World Cup matchSwitching Fitness Trackers in 2026 Can Preserve Old Health Data, but Cross-Brand Moves Need Manual ExportsTving Tops September Streaming App Installations Despite Data Breach ControversySamsung, SK hynix to report Q3 earnings with AI memory demand in focus
Large Language Models

Tutorial Tests Underdog Saluki 27B's 2-Bit Tool-Calling Claim Against ISTA Quantization

A MarkTechPost tutorial shows how to run Underdog Saluki 27B, a 7.9 GB two-bit Qwen3.8-27B build, and benchmark its tool-calling claims against the ISTA-DASLab model it is based on.

AI News Feed · 2026-10-25 13:12

Saluki is based on ISTA-DASLab's GSQ-RCO quantization of the same Qwen3.8-27B model. The walkthrough places both GGUF files side by side in stock llama.cpp. It starts with a tensor-by-tensor diff of the two GGUF headers and a byte-level comparison intended to show whether Saluki is a remix of ISTA's bits or a changed model.

The tutorial then serves each file on one GPU, reads its memory use and speed, and scores both on the same seeded 120 tasks from the Berkeley Function Calling Leaderboard. It uses BFCL's own checker with strict and lenient parsing. The BFCL v4 material covers simple Python, multiple, irrelevance, parallel and parallel multiple tasks. The notebook records each step's result and tracks each maker claim as held, false or undecided.

The setup is pinned to llama.cpp release b11565, dated Oct. 12, 2026, with commit f9a95aeb2. It pins the Saluki model to the ConwayResearch/Underdog-Saluki-27B-1.0 repository and a 7,898,369,152-byte GGUF file with SHA-256 hash 4a673518b11b1c4445f9b9a9d3c40356f5ba39f6a6260dcb25f3fbfd475d9efb. The ISTA file is pinned to ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF and a 8,422,841,472-byte GGUF with SHA-256 hash f0ae5006da0ce6225935339e4e989369f94de95d2263cf969519f8420c9ae02c. BFCL v4 data and checker files are taken from a single gorilla commit.

The tutorial states that it needs an NVIDIA GPU on Linux, including a Colab T4, or an Apple-silicon Mac. Everything the notebook downloads is pinned by size and SHA-256, including the llama.cpp binary, model files, BFCL data and checker code. MarkTechPost presents the 2-bit tool-calling result and the comparison with the 54 GB model as the makers' claims, with the notebook built to run the head-to-head test.

Editor's Summary

The MarkTechPost tutorial shows how to run and benchmark Underdog Saluki 27B, a 7.9 GB two-bit Qwen3.8-27B build, against ISTA-DASLab's GSQ-RCO version. It pins llama.cpp, model hashes and BFCL v4 data to compare tool-calling accuracy, memory and speed. The maker's claim that Saluki preserves tool calling and beats the full 54 GB model is tested through a seeded 120-task BFCL run.