73.3decode tok/s
1,553prefill tok/s
32Kcontext
Tested on this card · Oct 9, 2026
Underdog Saluki 27B 1.0
ConwayResearch's mixed 2-bit GGUF quantization of Qwen3.8-27B, using GSQ-RCO weights and an importance matrix for tool calling. This recipe serves text; the optional vision projector is not included.
1 hardware entries1 hardware configurations1 exact recipes
Hardware configurations
Each card is a hardware setup, including its card count, machine count and documented host requirements. The recipe dropdown updates measured speeds, context and the exact one-command launch together.
Speeds are the selected recipe’s recorded measurements, not estimates; “—” means unmeasured. Evidence distinguishes checks on this card, sibling-card checks, earlier checks and publisher reports.
Hugging Face model card
Underdog Saluki 27B 1.0
Qwen3.8-27B in under 8 GB, tuned to keep tool calling intact. A standard GGUF for stock llama.cpp. Read more: underdog.ai/saluki
Try Saluki today in underdog.ai
| Tool calling (Underdog Bench) | Parallel tool calls | Average retention | Size |
|---|---|---|---|
| 88 vs 84 for the full model | 42 vs 35 for the full model | 96% across 9 benchmarks | 7.89 GB vs 54 GB |
| Base model | Qwen3.8-27B |
| File | Underdog-Saluki-27B-1.0-IQ2-mix.gguf, 7.89 GB |
| Runtime | stock llama.cpp, and apps built on it |
| Modality | text, plus images with the optional vision add-on |
| Vision add-on | mmproj-Underdog-Saluki-27B-1.0-F16.gguf (928 MB) or -Q8_0.gguf (629 MB) |
| Thinking | on by default, switchable per request, same as Qwen3.8 |
Quickstart
huggingface-cli download ConwayResearch/Underdog-Saluki-27B-1.0 Underdog-Saluki-27B-1.0-IQ2-mix.gguf --local-dir .
llama-server -m Underdog-Saluki-27B-1.0-IQ2-mix.gguf --jinja -ngl 99 -fa on -c 32768
Add vision. Download the vision add-on too, and pass it with --mmproj. Then send images through the OpenAI chat API as usual.
huggingface-cli download ConwayResearch/Underdog-Saluki-27B-1.0 mmproj-Underdog-Saluki-27B-1.0-F16.gguf --local-dir .
llama-server -m Underdog-Saluki-27B-1.0-IQ2-mix.gguf --mmproj mmproj-Underdog-Saluki-27B-1.0-F16.gguf --jinja -ngl 99 -fa on -c 32768
--jinja turns on the Qwen3.8 chat template, which handles tool calls and thinking. The server speaks the OpenAI chat API on port 8080.
| Use | Settings |
|---|---|
| General use, reasoning, instructions | thinking on (default), temperature 0.6, top_p 0.95, top_k 20 |
| Fast, direct tool calls | thinking off with "chat_template_kwargs": {"enable_thinking": false}, temperature 0 |
Benchmarks
Saluki ran in stock llama.cpp. Where we ran the full-size Qwen3.8-27B ourselves, it used the same harness and settings.
Tool calling. Underdog Bench is 120 tasks from the Berkeley Function Calling Leaderboard (BFCL v4), frozen before any model was tested. Thinking off, temperature 0.
| Model | Size | Passed (of 120) |
|---|---|---|
| Underdog Saluki 27B 1.0 | 7.89 GB | 88 |
| Qwen3.8-27B, full size | 54 GB | 84 |
| Bonsai 2 | 5.95 GB | 70 |
Parallel tool calls. 100 BFCL v4 parallel tasks, official checker, thinking off: Saluki 42, full size 35.
Agents, instructions, coding and reasoning.
| Benchmark | Saluki | Full size |
|---|---|---|
| SWE-bench Verified (50 issues fixed) | 30 | 33 |
| IFEval (prompt-loose) | 93.5 | 91.5 (public) |
| IFBench (prompt-loose) | 72.7 | 71.0 (public) |
| MBPP+ | 78.0 | 83.9 (public) |
| MuSR | 67.5 | 79.6 (public) |
| AIME 2025 (avg@4) | 79.2 | 96.7 (public) |
| AIME 2026 (avg@4) | 80.0 | 94.6 (public) |
Public scores come from a different harness than ours. 120 tasks is a modest test, and a few tasks of any gap is run-to-run variation.
Limitations
- Keeps about 82 to 85% of the full model's competition math score.
- Weakest on letter-level instruction puzzles (palindromes, vowel rules, alphabetical order).
- About a fifth of parallel-call replies have small formatting slips.
- With thinking on, it often reasons at length before answering.
- Vision is an add-on. The main file is text only; add the 0.6 to 0.9 GB vision file when you need images.
Credits and license
Apache 2.0. Built on Qwen3.8-27B by the Qwen team and Qwen3.8-27B-GSQ-RCO-GGUF by ISTA-DASLab, both Apache 2.0. Underdog Bench tasks come from the Berkeley Function Calling Leaderboard. See NOTICE.
Built by Underdog.