base_model: Rootkit7/Laguna-S-2.1-uncensored
license: openmdw-1.1
tags:

• abliteration
• uncensored
• moe
• gguf
• llama.cpp

Laguna-S-2.1-uncensored — GGUF

GGUF quantizations of Rootkit7/Laguna-S-2.1-uncensored
— an uncensored (safety-refusal-removed) build of poolside's Laguna-S-2.1, a Mixture-of-Experts
reasoning model. The refusals are removed while capability is preserved: on a held-out multilingual harmful
set, refusal drops from ~95% → 3% and Solutus's automatic capability gate passes (no measurable
degradation). Built with the Solutus abliteration toolkit — the exact recipe is under Provenance & safety below.

⚠️ Runtime requirement — read this first. Laguna's hybrid architecture needs llama.cpp
with Laguna support: poolside's fork (git clone --branch laguna https://github.com/poolsideai/llama.cpp) or a recent-enough upstream (support is in
ggml-org/llama.cpp #25165, ~release b10087). Older llama.cpp — and current Ollama /
LM Studio — will not load these files yet.

ℹ️ Measured & gate-verified. Laguna is now a whitelisted architecture in Solutus; these numbers passed
the honest capability gate and every quant here was run-validated (see below). They are measured on
Laguna-S-2.1 specifically — indicative for that model, not a cross-model certification.

✅ Validation status — all quants run-validated. Q4_K_M, Q5_K_M, Q6_K, and Q8_0 were each loaded on the
poolside llama.cpp@laguna fork and checked live: they load and serve, comply on a harmful prompt
(abliteration carried through), stay coherent on a benign prompt, and reasoning works (a separate
reasoning_content thinking trace plus the correct final answer). For full precision, use the safetensors
repo (linked below).

Quants

All quants (Q4_K_M / Q5_K_M / Q6_K / Q8_0) are run-validated — see the validation note above. For
full precision, use the safetensors model at
Rootkit7/Laguna-S-2.1-uncensored.

Run

git clone --branch laguna https://github.com/poolsideai/llama.cpp
cd llama.cpp && cmake -B build -DGGML_CUDA=ON && cmake --build build -j --target llama-cli

# chat (reasoning on by default — this is a thinking model; use sampling, not greedy):
./build/bin/llama-cli -m Laguna-S-2.1-uncensored-Q4_K_M.gguf --jinja --temp 0.7
# reasoning off for a direct answer:
./build/bin/llama-cli -m Laguna-S-2.1-uncensored-Q4_K_M.gguf --jinja -rea off --temp 0.7 -p "..."

Provenance & safety

• Base: poolside/Laguna-S-2.1 @ 00af5a51782109b587a3b3bbf11875e566036fa7; recipe: Solutus ega,
  union of 8 datasets, α=5, router_scale=0.74.
• Safety refusals removed — a research artifact. It will comply with harmful requests. Use responsibly
  and per the base model's license, OpenMDW-1.1 (inherited from poolside/Laguna-S-2.1; a permissive
  open-weights license allowing use, modification, and redistribution incl. derivatives).