Laguna-S-2.1 — PS-IQ2_XXS
Compact GGUF quantization of
Laguna-S-2.1, released by PeasantSmith. Inspired by Antirez's mixed-precision GGUF work.
Base model
MoE, 118B total / 8B active params: 256 routed experts + 1 shared, top-10 routing, 48 layers (12 global attention + 36 sliding-window, window 512), GQA. Context up to 1,048,576 tokens via YaRN scaling on the global layers. Built for agentic coding and long-horizon tasks.
Quality
| Model | PPL |
|---|
| Laguna-S-2.1 F16 reference | 9.50 |
| Laguna-S-2.1 PS-IQ2_XXS | 11.5095 ± 0.26229 |
- Needle retrieval: PASS
- LRU-cache implementation: PASS
- File size: 33.65 GB, ~2.28 bits/param
PPL was measured with llama-perplexity on the evaluation corpus. Results are corpus-dependent.
Recommended settings
1temperature: 1.0
2top_k: 20
3top_p: 1.0
Compatible with GGUF runtimes including llama.cpp and Ollama.
License
OpenMDW-1.1, inherited from the base model.
Notes
Independently produced, experimental release with limited compute for testing. Feedback on quality, stability, and cross-runtime compatibility is welcome. See the upstream model page for full docs and license terms; verify behavior for your workload before relying on this in production.
