grug-v1.1-qwen-3.8-27b — GGUF
grug in box for llama.cpp. same brain as ProCreations/grug-v1.1-qwen-3.8-27b, just smaller box.
| file | size | who for |
|---|---|---|
grug-27b-v1.1-Q8_0.gguf |
28.6 GB | want closest to full weight |
grug-27b-v1.1-Q6_K.gguf |
22.1 GB | very good, less space |
grug-27b-v1.1-Q5_K_M.gguf |
19.2 GB | good middle |
grug-27b-v1.1-Q4_K_M.gguf |
16.5 GB | grug pick this one for most cave |
grug-27b-v1.1-Q3_K_M.gguf |
13.3 GB | small cave, tight RAM |
mmproj-grug-27b-v1.1-f16.gguf |
0.9 GB | eyes. only need if you show picture |
every quant load-tested with llama-bench before upload. built with llama.cpp
7c35571e.
run grug
llama-cli -m grug-27b-v1.1-Q4_K_M.gguf -p "write a function that flattens a nested list"
with eyes:
llama-mtmd-cli -m grug-27b-v1.1-Q4_K_M.gguf \
--mmproj mmproj-grug-27b-v1.1-f16.gguf --image cave-painting.png -p "what this?"
grug think small
grug reason inside <think> in caveman, then answer in normal english. on agent
step grug spend 20 think token where base model spend 108.5. on HumanEval
grug spend 79.5 where base spend 559.
use medium reasoning effort. xhigh make grug worse at picking tool (76.5 vs 97.1) — full table and chart on the main model card.
apache-2.0.