Laguna XS 2.1, self-quantized to GGUF by Atomic Chat. Built straight from Poolside's original weights with a per-tensor importance matrix, so this is not a repack of somebody else's files. Runs fully offline.
Highlights
- 33.4B parameters: the weights this repo quantizes.
- Context length: 262,144 tokens (256K), as published by Poolside.
- 40 layers: Mixture-of-Experts, hybrid sliding-window (512) and global attention.
- Full imatrix ladder: every quant is calibrated with an importance matrix.
- Mixed SWA and global attention layout: Laguna XS 2.1 uses sigmoid gating with per-layer rotary scales, enabling mixed SWA (Sliding Window Attention) and global attention layers in a 3:1 ratio (across 40 total layers).
- KV cache in FP8: KV cache quantized to FP8, reducing memory per token.
- Native reasoning support: Interleaved thinking between tool calls with support for enabling and disabling thinking per-request.
[!NOTE] These GGUFs are self-quantized from the original weights, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.
[!IMPORTANT] Always pass
--jinjaso the Laguna XS 2.1 chat template is applied. Without it the model can emit malformed turns.
Model Overview
| Property | Value |
|---|---|
| Base model | poolside/Laguna-XS-2.1 |
| Parameters | 33.4B |
| Layers | 40 |
| Experts | 256 routed (top-8) |
| Sliding window | 512 tokens |
| Context length | 262,144 tokens (256K) |
| Vocabulary | 100,352 |
| Modalities | Text |
| Architecture | Mixture-of-Experts, 256 experts (top-8), hybrid sliding-window (512) and global attention, 48 attention heads over 8 KV heads, LagunaForCausalLM |
| This repo | GGUF quants (imatrix). Quants: Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0 |
Scores are Poolside's published results for the base poolside/Laguna-XS-2.1, not our own measurements. Quantization preserves the large majority of this; Q4_K_M and up stay close to full precision.
Choosing a quant
| Quant | Size | Notes |
|---|---|---|
Q3_K_M | 16.1 GB | Low quality but usable. |
Q4_K_M | 20.3 GB | Recommended default. Best balance of size, speed and quality. |
Q5_K_M | 23.8 GB | Higher quality, low loss. |
Q6_K | 27.5 GB | Near lossless, noticeably lighter than Q8_0. |
Q8_0 | 35.6 GB | Effectively lossless, reference quality. |
[!TIP] Pick the largest file that fits your (V)RAM with room for context.
Q4_K_Mis the sweet spot for most setups;Q6_KorQ8_0for maximum fidelity.
Get started
Run Laguna XS 2.1 locally with:
- Atomic Chat: the easiest path. Open the app, search
AtomicChat/Laguna-XS-2.1-GGUF, pick a quant, hit Use this model. - llama.cpp:
llama-server -hf AtomicChat/Laguna-XS-2.1-GGUF:Q4_K_M --jinja -c 8192 - Ollama:
ollama run hf.co/AtomicChat/Laguna-XS-2.1-GGUF:Q4_K_M - LM Studio / Jan: search the repo id, download any quant.
Best practices
| Parameter | Value |
|---|---|
| temperature | 1.0 |
| top_p | 1 |
| top_k | 20 |
| min_p | 0.0 |
Poolside's recommended sampling configuration for poolside/Laguna-XS-2.1.
Run in llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server
./llama.cpp/build/bin/llama-server \
-hf AtomicChat/Laguna-XS-2.1-GGUF:Q4_K_M \
--jinja -ngl 99 -c 8192 -fa on
How these were made
- Download
poolside/Laguna-XS-2.1(original weights). - Convert to f16 GGUF with llama.cpp.
- Build an importance matrix over our calibration corpus.
- Quantize the ladder with
--imatrix.
License
Original model by Poolside, released under the OpenMDW-1.1 license. Full terms: OpenMDW-1.1. Quantized by Atomic Chat.