Laguna-XS-2.1 GGUF
Imatrix quants of poolside/Laguna-XS-2.1 — a 33B MoE coder with 3B active parameters and a 262k context.
Eight tiers in two families: UDX spends its bits on the tensors every token touches, APEX trades some of that for a smaller file.
| file | size | active bpw | file bpw |
|---|---|---|---|
UDX-Q6_K |
28.12 GB | 8.01 | 6.73 |
APEX-I-Balanced-v2D-lite |
24.37 GB | 6.75 | 5.83 |
UDX-Q5_K_S |
23.89 GB | 7.63 | 5.71 |
APEX-I-Quality-v2D-lite |
21.88 GB | 6.53 | 5.23 |
UDX-Q4_K_XL |
21.34 GB | 7.41 | 5.10 |
UDX-IQ4_XS |
16.95 GB | 6.88 | 4.05 |
APEX-I-Compact-v2D-lite |
15.84 GB | 4.87 | 3.79 |
APEX-I-Mini-v2D-lite |
12.89 GB | 4.06 | 3.05 |
Which one to pick
Use the active bpw column, not file size.
This model routes 8 of its 256 experts per token, so the expert weights — about 97% of the file — only contribute 8/256 of their precision to any given token. Attention and the shared expert are used on every token. Active bpw measures what the model actually runs at; file bpw measures what it costs on disk.
That's why UDX-IQ4_XS (16.95 GB) runs at 6.88 active bpw while
APEX-I-Compact-v2D-lite (15.84 GB) runs at 4.87 — nearly the same download, a
meaningfully different model.
- Best quality that still fits 24 GB VRAM —
UDX-Q4_K_XL - Best quality per GB —
UDX-IQ4_XS - Smallest usable —
APEX-I-Mini-v2D-lite - Near-lossless —
UDX-Q6_K
Run it
llama-server -m Laguna-XS-2.1-UDX-Q4_K_XL.gguf --jinja -ngl 99 -fa on
Needs llama.cpp release b10087 or newer (the base model card still says
Laguna support is unmerged upstream — it landed in
#25165 on 2026-07-22).
Context is 262,144. To go higher:
--ctx-size 1048576 --rope-scaling yarn --rope-scale 128 --yarn-orig-ctx 8192
Speculative decoding: poolside ship a separate DFlash drafter, but it currently needs poolside's llama.cpp fork for both the drafter and the server. These files work normally without it.
Notes
Imatrix by bartowski, computed on Laguna-XS-2.1 itself (437k tokens). Applied to every tier here.
APEX tiers replay mudler's per-layer maps, with attention lifted on the 10 global-attention layers — the ones carrying long-range context, since the other 30 use a 512-token sliding window.
UDX tiers are named that way on purpose. Unsloth never published a UD ladder for this model; these follow the same allocation policy but were not built by Unsloth and shouldn't be read as their release.
Every file was verified by checksum after upload — see sha256sums.txt.