fraQtl/Gemma-4-E4B-it-Hi-Fi-GGUF

Gemma-4-E4B-it Q4KM GGUF — fraQtl Hi-Fi (calibration-aware quantization)

Calibration-aware Q4KM build of google/gemma-4-E4B-it. A modest but real fidelity improvement over the community Q4KM, strongest on code/math. Every number below is measured, 3 independent runs (run-to-run drift 0.0), and reproducible from the receipts in this repo.

Fidelity to the ORIGINAL bf16 model (KLD, lower = closer)

Honest read — stated plainly, not buried: the win is real on code/math (−16.2%), near-parity on general (−4.2%), and top-1 agreement is a tie (0.948 / 0.910 both). This is a smaller lift than our Gemma-4 E2B build (−36% / −27%) — E4B's larger footprint leaves less quantization headroom. If your workload is code/math-heavy on-device, this helps; on general chat it's about even with the community build.

Also included

• —Phone build (...-Phone.gguf): ~4.59 GB (~15% smaller), PLE demoted to Q4_K, speed-neutral by construction (bandwidth-relevant bytes identical).
• —GSM8K (0-shot instructed chain-of-thought, n=200, seed 0): comparable to the community build (bartowski 84.0%). No win claim.
• —Wikitext-2 perplexity: sanity flag only, not a gate.

Reproducibility

Gates 1/2 PASS + phone smoke PASS. Fidelity metric: symmetric top-20 KL divergence on a Q8 teacher of the ORIGINAL bf16 checkpoint (the axis a user migrating from bf16 experiences). The imatrix, recipe hash, and every artifact sha256 ship in this repo. Runs via llama.cpp / the GenieX llama.cpp plugin (on-device / edge: Snapdragon X-class laptops, phones).

Built by fraQtl. Community builders credited: bartowski Q4KM baseline.