llama-cli -hf Chungulus/Agents-A1-IQ4_XS-imatrix-gguf-fable5-calibrated:Agents-A1-IQ4_XS-imatrix-gguf-fable5-calibrated.gguf -p "Write a Python sorting function" -n 160

Ollama

bash

ollama create agents-a1-iq4_xs-imatrix -f Modelfile
ollama run agents-a1-iq4_xs-imatrix

Which File Should I Download?

Use case	Recommendation
Recommended hardware	16-24 GB RAM
Best for	efficient compact 4-bit

Quality Snapshot

F16 baseline mini accuracy: 89.58%. F16 baseline PPL on KL holdout: 13.0194.

Metric	Value
Mini accuracy	87.50%
Retention vs F16	97.67%
Mean KLD vs F16	0.020185
Same top p	92.06%

Notes

Calibration source: Glint-Research/Fable-5-traces
Calibration source revision: e05c417852fc59fd8da758e68b352732423ca0cb
GGUF quantization method: llama.cpp with imatrix calibration.
Static imatrix GGUF; not Unsloth Dynamic 2.0 / UD2.
MTP is not included because the downloaded checkpoint did not contain MTP tensors.
This repo contains local quantization artifacts only.