---
base_model:
- MiniMaxAI/MiniMax-M3
---

# Experimental ik_llama.cpp quant pipeline

⚠️ These are untested artifacts from an experimental ik_llama.cpp quant pipeline


## PPL on wiki.raw

This IQ2_KT quant (110.9 GiB):

Final estimate: PPL over 552 chunks for n_ctx=512 = 7.5871 +/- 0.05498

The IQ3_KT quant (156.9 GiB):

Final estimate: PPL over 552 chunks for n_ctx=512 = 6.0129 +/- 0.04200

Unsloth UD_Q4_K_M (246.7 GiB):

Final estimate: PPL over 552 chunks for n_ctx=512 = 5.2593 +/- 0.03521