I host 30+ free APEX MoE quantizations as independent research. My only local hardware is an NVIDIA DGX Spark (122 GB unified memory), enough for ~30-50B-class MoEs, but bigger ones (200B+) require rented compute on H100/H200/Blackwell, typically $20-100 per quant. If APEX quants are useful to you, your support directly funds those bigger runs.
APEX (Adaptive Precision for EXpert Models) quantizations of poolside/Laguna-XS-2.1 — poolside's Laguna XS.2 Mixture-of-Experts model for coding and agentic software engineering.
Requires a recent llama.cpp with Laguna support (PR #25165). Older builds cannot load arch=laguna.
Available Files
File
Profile
Best For
Laguna-XS-2.1-APEX-I-Balanced.gguf
I-Balanced
Best overall — imatrix-enhanced
Laguna-XS-2.1-APEX-I-Quality.gguf
I-Quality
Highest quality with imatrix
Laguna-XS-2.1-APEX-Quality.gguf
Quality
Highest quality (no imatrix)
Laguna-XS-2.1-APEX-Balanced.gguf
Balanced
General purpose
Laguna-XS-2.1-APEX-I-Compact.gguf
I-Compact
Consumer GPUs, imatrix-enhanced
Laguna-XS-2.1-APEX-Compact.gguf
Compact
Consumer GPUs
Laguna-XS-2.1-APEX-I-Mini.gguf
I-Mini
Smallest viable, fastest inference
What is APEX?
APEX is a quantization strategy for Mixture-of-Experts (MoE) models. It classifies tensors by role (routed expert, shared expert, attention, dense FFN) and applies a layer-wise precision gradient — edge layers get higher precision, middle layers compress more aggressively. I-variants use diverse imatrix calibration (chat, code, reasoning, tool-calling, agentic traces, Wikipedia).
In MoE models the routed-expert FFN tensors dominate the weight budget but only ~8/256 experts fire per token, so APEX compresses middle-layer routed experts hardest while preserving edge layers, attention, and the always-active shared expert.
APEX layout for Laguna
Laguna XS.2 has a structure APEX handles explicitly:
Layer 0 is a leading dense FFN (no experts) — pinned to Q8_0, since every token traverses it.
Layers 1–39 are MoE — 256 routed experts + a shared expert, 8 active per token, sigmoid gating.
Shared expert (ffn_*_shexp) kept at Q8_0 on every tier (always active).
Routed experts follow the 5+5 symmetric edge gradient (higher precision at the first/last layers, most aggressive in the middle).
Router (ffn_gate_inp), norms and the exp_probs_b gating bias stay at full precision.
