🌊 About this release

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:

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.
APEX mixed-precision GGUF from a poolside llama.cpp (laguna) BF16 GGUF, using APEX-style tensor-type configs + imatrix (token embedding / output kept at BF16).

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

⚠️ Uncensored notice

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.

Refusals (harmful eval)	8 / 100 (baseline ~97 / 100)
KL divergence	~0.0042
Length deviation	~0.07 σ
Generation health	PASSED
Selected trial	abliterix Trial 16

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 metadata).

📜 Disclaimer (identity & text quality)

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. More aggressive quant tiers (e.g. Compact) 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.

🧠 Model details

Architecture	Laguna MoE (poolside `laguna`)
Parameters	~118B total, ~8B active / token
Layers	48 (L0 dense FFN, L1–L47 MoE)
Experts	256 routed + 1 shared, top-10
Attention	GQA, 8 KV heads, head dim 128
Context	Up to ~1,048,576 tokens (practical limit depends on VRAM / `-c`)
Vocab	100,352 (Laguna family tokenizer)
Modality	Text → text (no mmproj in this package)
This repo	APEX I-Quality / I-Balanced / I-Compact GGUF

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

💡 What is APEX?

These files use APEX-style MoE-aware mixed precision: precision follows tensor role + layer position (higher on edges, more aggressive in the middle), with imatrix for the I- tiers.

Common settings for this package: source Laguna-S-2.1-Uncensored BF16 GGUF; calibration laguna-s-2.1.imatrix; token embedding / output = BF16; routers and norms largely left at high precision defaults. Configs target Laguna 48-layer naming (ffn_*_exps / ffn_*_shexp / attn_*, L0 dense).

📦 APEX quantization tiers

File	Size	Mid experts	Best for
`*-I-Quality.gguf`	~70 GB	edge Q6_K / near Q5_K / mid iq4_xs; shared Q8_0; attn Q6_K	Smaller high-quality try (IQ mid-layers)
`*-I-Balanced.gguf`	~80 GB	edge Q6_K / near & mid Q5_K; shared Q8_0; attn Q6_K	Recommended default for Chinese users — steadier than Compact
`*-I-Compact.gguf`	~52 GB	edge Q4_K / mid Q3_K; shared Q6_K; attn Q4_K	Tighter memory; more quant noise OK

How to choose:

Chinese-heavy use: prefer I-Balanced. Mid experts stay Q5_K (not iq4_xs), which usually feels more stable. Drop to I-Compact only if memory is tight.
English / coding: tier differences are usually small; pick by VRAM/speed (Balanced vs Compact). I-Quality is optional when you want IQ mid-layers and a smaller footprint than Balanced.

🚀 Usage (llama.cpp)

Use a build that understands the laguna architecture (poolside laguna branch or equivalent).

Example (I-Balanced recommended)

./llama-server \
  -m ./Laguna-S-2.1-Uncensored-APEX-I-Balanced.gguf \
  --port 8080 \
  -sm none \
  --device rocm0 \
  --ctx-size 131072 \
  --flash-attn on \
  --no-mmap \
  --fit on \
  --jinja \
  --host 0.0.0.0

Must recognize general.architecture = laguna.
Thinking / tools: configure per client/server docs (e.g. enable_thinking, reasoning options).
Warnings like special_eos_id is not in special_eog_ids are tokenizer metadata hints; the model can still load. If stop/truncation is odd, check stop / max_tokens / template.
Text-only GGUF; no mmproj.

🎛️ Recommended sampling

General / coding	temperature 0.6–1.0, top_p 0.95, top_k 20
More stable	Slightly lower temperature

🔧 Build pipeline (summary)

Base: poolside/Laguna-S-2.1
abliterix search → Trial 16 LoRA + MoE router adjustments
BF16 stream-merge → Laguna-S-2.1-Uncensored
poolside llama.cpp → BF16 GGUF + imatrix
APEX tensor-type-file + imatrix → I-Quality / I-Balanced / I-Compact

Laguna-S-2.1-Uncensored-APEX-GGUF