This repository contains a reduced-memory Laguna S 2.1 quantization for
DwarfStar.
Mixed Q2_K/Q3_K variant
laguna-s-2.1-RoutedQ2_K-Last27Q3_K.gguf keeps every non-routed tensor
byte-identical to Poolside's laguna-s-2.1-Q4_K_M.gguf at revision
706fa69799926b6afde1af9e24ca2a4923f110a1. Only routed expert tensors were
requantized using the source importance matrix:
routed layers 1 through 20: Q2_K gate, up, and down
routed layers 21 through 47: Q3_K gate, up, and down
all other tensors: unchanged from the official Q4_K_M GGUF
The file is 48,260,803,968 bytes (44.946 GiB), intended for full-residency
inference on 64 GiB systems. Runtime memory also depends on context size and
KV-cache allocation.
On an Apple M5 Max, the tested model reached approximately 514 tokens/second
for a 4096-token prefill and 63 tokens/second steady-state generation.
Against 100 official continuation vectors, the mixed model obtained average
NLL 0.2583, 87/100 first-token matches, and average matching-prefix length
9.50 tokens. The corresponding full Q4_K_M measurements were 0.2352, 92/100,
and 10.86.
