0xSero / Sybil Solutions

Gemma REAP Family

Gemma family REAP releases and calibration artifacts.

donate.sybilsolutions.ai REAP paper REAP repo PR17 implementation Cerebras REAP collection

What this profile is doing

This profile publishes live Mixture-of-Experts compression work: observer datasets, REAP-pruned checkpoints, quantized releases, and benchmarks around keeping quality while removing experts aggressively.

Models

0xSero/gemma-4-19b-a4b-it-REAP
0xSero/gemma-4-21b-a4b-it-REAP
0xSero/gemma-moe-reap

Calibration Datasets

0xSero/gemma-moe-reap

REAP

Router-weighted Expert Activation Pruning ranks experts by how much they contribute to routed computation, then removes low-saliency experts instead of merging everything blindly. This surface groups the family-specific artifacts so the compression story is obvious from the Hub profile.