OPEN LINK ↗
// 1h agoMODEL RELEASE
Underdog Saluki Shrinks Qwen3.8 for Tools
Underdog Saluki 27B compresses Qwen3.8-27B into a 7.89GB IQ2-mix GGUF tuned for local tool calling. It runs with stock llama.cpp and retains strong agent behavior, though math, reasoning, and coding performance decline. [Model card](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0)
// ANALYSIS
Saluki is a compelling local-agent tradeoff, but its tool-calling edge should be treated as promising rather than definitive.
- –The 7.89GB footprint makes a 27B-class model practical on consumer GPUs and local machines.
- –It scored 88/120 tool-calling tasks versus 84 for full Qwen3.8-27B, plus 42 versus 35 on parallel calls.
- –The benchmark is creator-run and modest in size, so the apparent win needs independent validation.
- –Math, reasoning, MBPP+, and SWE-bench results trail the full model, making it better suited to orchestration than deep problem-solving.
- –Apache 2.0 licensing and standard llama.cpp compatibility lower adoption friction for self-hosted agents.
// TAGS
underdog-saluki-27bllmopen-weightsquantizationtool-useinferencelocal-firstopen-source
DISCOVERED
1h ago
2026-10-24
PUBLISHED
1h ago
2026-10-24
RELEVANCE
10/ 10
AUTHOR
Better Stack