license: openmdw-1.1 base_model: poolside/Laguna-S-2.1 base_model_relation: quantized library_name: llama.cpp pipeline_tag: text-generation language:
- en
- de tags:
- laguna
- moe
- uncensored
- gguf
- quantized
- bilingual
- code
Laguna-S-2.1-Uncensored GGUF
GGUF quantization of ressl/Laguna-S-2.1-Uncensored, derived from poolside/Laguna-S-2.1.
Original uncensoring work by Robert Ressl — full model card and evaluation details are at the source repository.
Quantized Variants
All variants use the oscar llama.cpp build (v4310aa4f8) with native LLM_ARCH_LAGUNA support. The BF16 baseline was converted with convert_hf_to_gguf.py from the oscar repo, which handles Laguna's per-layer alternating head counts, sigmoid MoE routing, mixed full/sliding attention RoPE, and the shared-expert topology.
All quantizations use --leave-output-tensor to preserve output.weight in BF16 for generation quality.
| File | Size | Precision | Notes |
|---|---|---|---|
laguna-s-2.1-bf16-uncensored.gguf | 220 GiB | BF16 | Lossless round-trip from safetensors |
laguna-s-2.1-q8_0-uncensored.gguf | 117 GiB | Q8_0 uniform | All tensors Q8_0 |
laguna-s-2.1-q5_k_m-uncensored.gguf | 79 GiB | Q5_K_M uniform | All tensors Q5_K_M |
laguna-s-2.1-q4_k_m-uncensored.gguf | 67 GiB | Q4_K_M uniform | All tensors Q4_K_M |
laguna-s-2.1-uncensored-hq.gguf | 77 GiB | Mixed | Experts+Shared+Attn:Q5_K, Emb/Out/Routing:Q8_0 |
laguna-s-2.1-uncensored-balanced.gguf | 63 GiB | Mixed | Experts:Q4_K, Attn+Shared:Q5_K, Emb/Out/Routing:Q8_0 |
laguna-s-2.1-uncensored-compact.gguf | 49 GiB | Mixed | Experts:Q3_K, Attn+Shared+Dense:Q5_K, Emb/Out/Routing:Q8_0 |
laguna-s-2.1-uncensored-ultracompact.gguf | 49 GiB | Mixed | Experts:Q3_K, Attn:Q4_K, Shared+Dense:Q3_K, Emb/Out:Q4_K |
Mixed-precision recipes use per-tensor --tensor-type overrides for 814 tensors via --tensor-type-file. The default fallback type for unlisted tensors is the top-level quant type passed to llama-quantize.
Usage
llama-cli --model laguna-s-2.1-uncensored-balanced.gguf -ngl 0 -p "Hello" -n 128
For GPU offloading:
llama-cli --model laguna-s-2.1-uncensored-balanced.gguf -ngl 99 -c 32768
Model Details
| Architecture | Laguna MoE, 48 layers, 256 routed experts (top-10 sigmoid) + 1 shared expert |
| Parameters | 118B total, ~8B activated per token |
| Context length | 1,048,576 tokens (YaRN, rope base 500000 / 10000 SWA) |
| Vocab size | 100,352 |
| Languages | English, German |
| License | OpenMDW-1.1 |
Original Evaluation (ressl/Laguna-S-2.1-Uncensored)
| Metric | Base | Uncensored |
|---|---|---|
| English refusals (686 prompts) | 92.71% | 2.33% |
| German refusals (686 prompts) | 74.49% | 4.23% |
| XSTest over-refusal (214 prompts) | 8.88% | 1.87% |
| HumanEval pass@1 (164 problems) | 90.24% | 85.37% |
Full evaluation details, datasets, and methodology are in the source repository.
Recipe Selection Guide
| Use case | Recommended variant |
|---|---|
| Maximum quality | HQ or Q8_0 |
| Best size/quality tradeoff | Balanced |
| Lower VRAM, similar quality to balanced | Compact |
| Minimal disk footprint | Ultra Compact |
Credits
Original uncensoring: Robert Ressl (ressl.ch)
GGUF conversion and quantization: converted with oscar convert_hf_to_gguf.py (commit 4310aa4f8) and llama-quantize from the same build, using the registered LagunaForCausalLM conversion plugin.
GGUF collection generated August 2026.