Laguna S 2.1 is a poolside ~118B total / ~8B active per token Mixture-of-Experts model for agentic coding and long-horizon work. Architecture highlights: token-choice routing with softplus gates, 256 routed experts + 1 shared expert (top-10), GQA, 1:3 global/sliding-window attention (48 layers: 12 global + 36 local, window 512), and up to about 1M context, with optional native thinking interleaved with tool use.

This release builds on the official weights in two steps:

1. Uncensored behavior edit via abliterix (ROCm + bitsandbytes 4-bit search path), selecting Trial 16 LoRA, stream-merged back to BF16, plus MoE safety-expert router bake-in. Thinking was restored: the abliteration-time force-false override was removed from the chat template (see below).
2. Full-precision release: this repo ships the uncensored weights as BF16 safetensors (46 shards, ~219 GB). Quantized APEX I-tier GGUFs + imatrix built from this exact merge live in the sibling repo Laguna-S-2.1-Uncensored-APEX-GGUF.

License: OpenMDW-1.1 (same family as the base model). See openmdw.ai and poolside terms.

After merging abliterix Trial 16, this model shows a much lower refusal rate and can differ substantially from official Laguna-S-2.1. Evaluate compliance and safety for your use case; control access and audit as needed.

Implementation sketch: LoRA merge W += (B @ A) * (alpha / r) (this trial alpha = r = 1); per-layer top ~20 safety experts with router row scale ~0.74 (see MERGE_INFO.json).

The following can also appear with official Laguna-S-2.1 (including high-precision online deployments). They are not introduced by this Uncensored edit and should not be blamed on abliterix / the refusal merge:

• Model identity quirks or self-descriptions that diverge from the official persona;
• Odd tokens, typos, stiffness, or occasional awkward phrasing in Chinese and other languages.

This Uncensored pipeline mainly changes refusal / safety-related behavior. This repo ships full-precision BF16 weights, so no quantization noise is introduced here; if you use the APEX GGUFs, more aggressive tiers may amplify existing text noise, but the underlying issues already exist in the original model and/or the local quant+inference stack. For a fair check, compare official vs this release under the same serving settings.

Coding / tool ability is largely retained; refusal and alignment behavior are changed. Full official bench tables were not re-run for this derivative.

• model-00001-of-00046.safetensors … model-00046-of-00046.safetensors — BF16 full weights (46 shards, ~219 GB), stream-merged from poolside/Laguna-S-2.1 + Trial 16 LoRA + router bake-in.
• config.json / configuration_laguna.py / generation_config.json — laguna architecture; requires trust_remote_code=True. generation_config.json presets enable_thinking=true, poolside_v1 tool/reasoning parsers, and a DFlash speculative-decoding config.
• chat_template.jinja — thinking-restored template (enable_thinking | default(true)).
• MERGE_INFO.json — machine-readable provenance (trial, KL, refusals, router rows scaled, LoRA modules merged, audit count).
• UNCENSORED_README.md — condensed uncensored notes.
• LICENSE.md — OpenMDW-1.1.

The APEX-GGUF sibling repo is the quantization output of this exact merge: laguna-s-2.1.imatrix was calibrated on it, and token embedding / output stay BF16 there as well.

Transformers (trust remote code)

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"SC117/Laguna-S-2.1-Uncensored",
trust_remote_code=True,
torch_dtype="bfloat16",
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
"SC117/Laguna-S-2.1-Uncensored", trust_remote_code=True
)
messages = [{"role": "user", "content": "Implement a small HTTP server in Python."}]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(out[0], skip_special_tokens=True))

vLLM

vllm serve SC117/Laguna-S-2.1-Uncensored \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --tool-call-parser poolside_v1 \
    --reasoning-parser poolside_v1 \
    --enable-auto-tool-choice \
    --served-model-name laguna-uncensored \
    --default-chat-template-kwargs '{"enable_thinking": true}'

SGLang

python -m sglang.launch_server \
  --model-path SC117/Laguna-S-2.1-Uncensored \
  --tp-size 4 \
  --trust-remote-code \
  --reasoning-parser poolside_v1 \
  --tool-call-parser poolside_v1

• llama.cpp / GGUF users: use the APEX-GGUF repo with poolside's llama.cpp fork, branch laguna (must recognize general.architecture = laguna).
• Speculative decoding: pair with the DFlash draft model (dflash method, 15 draft tokens — already preset in generation_config.json).
• Thinking is controlled per request via the chat template: chat_template_kwargs={"enable_thinking": False} to disable.

1. Base: poolside/Laguna-S-2.1
2. abliterix search → Trial 16 LoRA + MoE router adjustments
3. BF16 stream-merge → this repo (Laguna-S-2.1-Uncensored, 46 shards)
4. GGUF branch: poolside llama.cpp → BF16 GGUF + imatrix → APEX tensor-type-file + imatrix → I-Quality / I-Balanced / I-Compact (see APEX-GGUF)