pearsonkyle/gemma4-31b-imatrix-mtp-GGUF

README: add static PPL/KLD/top_p to the imatrix-vs-AWQ tool-call table

Swap IQ2_M to AWQ build: +54% tool-argument accuracy and far tighter run-to-run variance vs plain 2-bit imatrix at the same 2.85 bpw. Static KLD/PPL slightly favored imatrix but mispredicted agentic tool-calling; AWQ chosen by a tool-call fidelity replay. README: add Method row, AWQ static column, tool-call table, how-it-was-made.

Ship Q4_K_M MTP drafter (zero acceptance penalty vs Q8_0, -156 MB); add drafter-quantization acceptance study + figure to README

README: correct Q5_K_S tool-error phrasing (on par with IQ4_XS)

README: fill Q5_K_S agentic results (40% pass / 100% patch)

README: replace Q5_K_M with Q5_K_S (24GB build), update MTP table

Add Q5_K_S imatrix quant (5.55 bpw) — text+vision+MTP under 24GB

Add Q5_K_M imatrix quant (5.69 bpw)

README: repoint self-links to renamed repo (gemma4-31b-imatrix-mtp-GGUF)

Remove MTP/ subdir copy (drafter moved to root)

README: MTP drafter moved to repo root; add -hf auto-discovery note

Move MTP drafter to repo root (mtp-gemma-4-31B-it.gguf) for -hf auto-discovery

README: MTP drafter section + acceptance-rate table (n=1..4 x quant)

Add Q8_0 MTP drafter (gemma4-assistant) for speculative decoding

Update README.md

README: vision/mmproj docs + worked example; links point to renamed repo

Add mecha.png vision demo asset

Add Q8_0 vision mmproj (text+image support)

Squash history: keep current models (IQ2_M/IQ3_M/IQ4_XS) + calibration data