Underdog Saluki 27B 1.0 puts a quantized Qwen3.8-27B checkpoint into a 7.89GB GGUF file for standard llama.cpp. Its creator reports 88 correct tool-calling tasks out of 120, against 84 for the full model. The useful question for a local assistant is which tasks changed—and how much memory the complete workload needs.
The released model card identifies an IQ2 mixed-precision build based on Qwen3.8-27B and ISTA-DASLab’s GSQ-RCO release. Apache 2.0 covers the checkpoint. This makes the file an accessible deployment option, while the benchmark evidence remains the developer’s evaluation rather than an EyesTech hardware test.
Four extra successes conceal twenty changed outcomes
Underdog’s launch breakdown supplies a more revealing number than the headline score: Saluki still passes 76 of the 84 tasks the original passed. Subtracting those overlaps gives eight previously successful tasks lost and twelve previously unsuccessful tasks gained. The net improvement is four tasks, or 3.33 percentage points; twenty of the 120 outcomes changed.
That matters when replacing a working assistant. An overall score increase can coexist with regressions on the exact functions a user depends on. The right comparison retains a per-task record: correct tool, correct arguments, unnecessary calls, omitted calls and parser failure. An average cannot tell a team whether its calendar or billing operation sits among those eight losses.
The evaluation uses a frozen subset of BFCL v4, with thinking disabled and temperature zero. The launch report itself warns that the set is modest and part of the gap may be run variation. It also separates task types: Saluki does well selecting one function, while repeated and parallel calls expose more friction. None of this establishes superiority across arbitrary agent workflows.
The parser changes the apparent model quality
On a separate 100-task parallel-call test, the creator reports 42 passes with strict parsing and 55 when small formatting differences are accepted. That is a thirteen-point change without changing the weights. The launch report also shows Mia ahead on this parallel subset, so the single-call headline should not determine a multi-action deployment.
A tolerant parser can normalize a known wrapper or equivalent representation. It should still reject an unknown function, missing required field or invalid argument value. Track recovered formatting errors separately from semantically correct calls; otherwise a parser improvement can be mistaken for a reasoning improvement. EyesTech’s FrogNano coverage describes another case where the surrounding tool interface materially changed a model’s score.
A 7.89GB file needs room around it
The weight file is approximately 6.84 times smaller than the developer’s 54GB reference. That is an 85.4% reduction in disk size, calculated from the two published sizes. It does not establish an 85.4% reduction in runtime memory or latency. Context state, compute buffers, the operating system and other applications still need capacity.
For image input, the current model files and card offer an additional vision projector: about 629MB for Q8 or 928MB for F16. Weight-plus-projector storage therefore rises to roughly 8.52GB or 8.82GB before runtime allocations. The older launch page says text only; the current downloadable release documents optional vision. Use the current card for that capability, and record both files in a reproducible setup.
The publisher’s llama.cpp recipe enables the Jinja chat template, GPU offload, flash attention and a 32,768-token context. These are configuration choices rather than proof that every 8GB GPU can hold the complete workload. Start with a smaller context and inspect the server’s allocation report, then increase prompt length and concurrency on the intended machine. Also record thinking mode: long reasoning traces can dominate response time even when weights fit comfortably.
Saluki is a practical candidate for compressed local tool use. Its strongest evidence is a usable standard-format release with explicit task-level caveats. A replacement decision should preserve the application’s existing successful cases, measure parser recovery separately and budget memory for the workload around the weights.
Prepared with AI assistance from the cited primary sources. Cover: AI-generated editorial illustration. This analysis uses published documentation and creator-reported evaluations; EyesTech did not run the model for this article.
