Underdog Saluki Packs Qwen3.8-27B into 7.89 GB for Local Tool Calling
A compact GGUF build of Qwen3.8-27B targets local agents; its creator reports strong tool-call results, while Hugging Face records more than 100,000 cumulative downloads.
A smaller file for local agents
Underdog’s Saluki 27B is a 7.89 GB GGUF package of Qwen3.8-27B that the creator positions for local tool calling. Its English model card says the main file runs in stock llama.cpp and keeps text generation separate from an optional vision add-on. The repository lists the model file and two downloadable vision files, making the packaging itself checkable without relying on a performance claim.
Underdog reports that Saluki passed 88 of 120 tasks in its tool-calling test, compared with 84 for the full-size model under its settings. It also reports 42 successful parallel-tool tasks out of 100, versus 35 for the full model. Those are the creator’s results, not independent replication. The card warns that 120 tasks is a modest sample, some comparisons use different test setups, and the quantized model loses more ground on competition math. Its claimed advantage should be read as a targeted tool-calling pitch rather than proof that a smaller file is universally stronger.
The current signal
Hugging Face dates the repository’s creation to October 8 and its latest model-card edit to October 10. In an October 12 trend snapshot, the platform showed 108,900 cumulative downloads and 342 likes. Those counters do not reveal distinct users or a download growth rate. The story is about a recently released community quantization drawing current attention, not a fresh model launch on October 12.