source: grug-27b (v2.1)
scheme: NVFP4 (W4A4) via llm-compressor one-shot
calibration: 512 samples of grug's own training distribution (agent sessions + grug reasoning) so scales match real usage
kept high precision: lm_head, vision tower, router gates

how run

bash

vllm serve ProCreations/grug-27b-nvfp4 --max-model-len 32768 \
  --reasoning-parser deepseek_r1

best on Blackwell (native FP4). works on Hopper too (weight-only benefit). grug think dense inside <think> (arrives in message.reasoning), answer normal english. sampling temp 0.6-1.0, top_p 0.95, top_k 20.

family: grug-27b | grug-27b-mtp | gguf | qat-q4

grug made by ProCreations.