Not-For-All-Audiences

Instructions to use dealignai/Laguna-S-2.1-CRACK-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="dealignai/Laguna-S-2.1-CRACK-GGUF",
	filename="Laguna-S-2.1-CRACK-Q2_K.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Use Docker

docker model run hf.co/dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dealignai/Laguna-S-2.1-CRACK-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dealignai/Laguna-S-2.1-CRACK-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Ollama
How to use dealignai/Laguna-S-2.1-CRACK-GGUF with Ollama:
```
ollama run hf.co/dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M
```

Unsloth Studio

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for dealignai/Laguna-S-2.1-CRACK-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for dealignai/Laguna-S-2.1-CRACK-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for dealignai/Laguna-S-2.1-CRACK-GGUF to start chatting

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use dealignai/Laguna-S-2.1-CRACK-GGUF with Docker Model Runner:
```
docker model run hf.co/dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M
```

Lemonade

How to use dealignai/Laguna-S-2.1-CRACK-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dealignai/Laguna-S-2.1-CRACK-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Laguna-S-2.1-CRACK-GGUF-Q4_K_M

List all available models

lemonade list

dealign.ai mascot

Laguna S 2.1 · CRACK · GGUF

118B-total / 8B-active MoE coding model · llama.cpp GGUF · Q6_K · Q4_K_M · Q2_K
Refusal removed with capability preserved — 0 refusals across HarmBench-320, MMLU within ~1.5 pts of base

A permissive research variant of poolside/Laguna-S-2.1 — the open-weight agentic coding Mixture-of-Experts model — packaged for llama.cpp as GGUF. Refusal behaviour has been removed while the model's coding, reasoning, tool-use, and general knowledge are preserved.

This public card intentionally describes compatibility, evaluation, and limitations only. Internal creation details are not published.

Model details

Property	Value
Base model	poolside/Laguna-S-2.1
Architecture	Laguna MoE — 118B total / ~8B active, 256 routed experts + 1 shared, top-10 router
Layers	48 (12 global : 36 sliding-window attention)
Context	up to 1,048,576 tokens · thinking and non-thinking modes
Format	GGUF for llama.cpp
Quant scheme	High-precision attention + router; low-bit routed experts (importance-matrix calibrated)
Languages	English, Chinese (+ the base model's full multilingual coverage)

The tokenizer, Laguna chat template (with <think> reasoning and tool-call support), and generation config are bundled. Requires a recent llama.cpp build with Laguna architecture support.

Quantizations

File	Bits	Size	Recommended for
`Laguna-S-2.1-CRACK-Q6_K.gguf`	~6-bit	~96 GB	Highest quality; near-lossless
`Laguna-S-2.1-CRACK-Q4_K_M.gguf`	~4-bit	~73 GB	Best balance — recommended default
`Laguna-S-2.1-CRACK-Q2_K.gguf`	~2-bit	~45 GB	Smallest; runs on less memory

All quants use importance-matrix calibration and keep attention and the router at high precision, so quality holds up even at 2-bit.

Evaluation

Capability is preserved — MMLU-logit accuracy vs the unmodified base model:

Quant	MMLU	Δ vs base
base (reference)	74.48%	—
Q6_K	72.75%	−1.73
Q4_K_M	74.27%	−0.21
Q2_K	73.67%	−0.81

_{MMLU-logit, full 14,042-question test, reasoning off. All quants sit well within the ±3% capability band of base — Q4_K_M is essentially lossless (−0.21).}

Per-topic MMLU: base vs CRACK — all 57 subjects (click to expand)

Full 14,042-question MMLU-logit, reasoning off. Δ = points vs base at the same measurement.

Summary by super-category

Category	Base	Q6_K	Q4_K_M	Q2_K
STEM	71.5	70.2 (−1.4)	71.5 (−0.0)	70.6 (−0.9)
Humanities	68.2	66.9 (−1.3)	67.8 (−0.3)	67.7 (−0.5)
Social Sciences	82.5	79.3 (−3.2)	82.0 (−0.5)	81.3 (−1.1)
Other (business, health, misc.)	78.8	77.5 (−1.4)	78.9 (+0.1)	77.9 (−0.9)
Overall	74.48	72.75 (−1.73)	74.27 (−0.21)	73.67 (−0.81)

All 57 subjects

Subject	Base	Q6_K	Q4_K_M	Q2_K
STEM
abstract algebra	60.0	59.0 (−1.0)	60.0 (+0.0)	60.0 (+0.0)
astronomy	86.2	87.5 (+1.3)	88.8 (+2.6)	87.5 (+1.3)
college biology	85.4	82.6 (−2.8)	85.4 (+0.0)	83.3 (−2.1)
college chemistry	50.0	52.0 (+2.0)	50.0 (+0.0)	52.0 (+2.0)
college computer science	72.0	69.0 (−3.0)	71.0 (−1.0)	72.0 (+0.0)
college mathematics	53.0	53.0 (+0.0)	54.0 (+1.0)	55.0 (+2.0)
college physics	61.8	59.8 (−2.0)	59.8 (−2.0)	57.8 (−3.9)
computer security	81.0	80.0 (−1.0)	81.0 (+0.0)	81.0 (+0.0)
conceptual physics	85.5	85.5 (+0.0)	86.8 (+1.3)	86.0 (+0.4)
electrical engineering	75.2	73.8 (−1.4)	75.9 (+0.7)	76.6 (+1.4)
elementary mathematics	66.7	65.3 (−1.3)	66.4 (−0.3)	64.3 (−2.4)
high school biology	87.7	85.8 (−1.9)	88.1 (+0.3)	87.1 (−0.6)
high school chemistry	71.4	69.0 (−2.5)	70.4 (−1.0)	69.0 (−2.5)
high school computer science	89.0	85.0 (−4.0)	87.0 (−2.0)	88.0 (−1.0)
high school mathematics	50.7	48.9 (−1.9)	50.0 (−0.7)	48.5 (−2.2)
high school physics	62.9	60.3 (−2.6)	62.9 (+0.0)	58.3 (−4.6)
high school statistics	73.6	71.3 (−2.3)	71.8 (−1.9)	74.5 (+0.9)
machine learning	59.8	61.6 (+1.8)	62.5 (+2.7)	58.0 (−1.8)
Humanities
formal logic	60.3	60.3 (+0.0)	61.1 (+0.8)	57.1 (−3.2)
high school european history	81.2	80.0 (−1.2)	79.4 (−1.8)	83.0 (+1.8)
high school us history	89.2	87.3 (−2.0)	87.7 (−1.5)	87.3 (−2.0)
high school world history	88.2	85.7 (−2.5)	88.2 (+0.0)	89.0 (+0.8)
international law	83.5	80.2 (−3.3)	85.1 (+1.7)	83.5 (+0.0)
jurisprudence	84.3	79.6 (−4.6)	82.4 (−1.9)	81.5 (−2.8)
logical fallacies	83.4	79.8 (−3.7)	83.4 (+0.0)	81.0 (−2.5)
moral disputes	74.9	74.9 (+0.0)	76.9 (+2.0)	73.7 (−1.2)
moral scenarios	52.1	50.8 (−1.2)	52.4 (+0.3)	57.0 (+4.9)
philosophy	79.1	78.1 (−1.0)	79.4 (+0.3)	77.5 (−1.6)
prehistory	84.6	83.6 (−0.9)	84.3 (−0.3)	82.7 (−1.9)
professional law	58.0	56.8 (−1.2)	56.4 (−1.6)	55.2 (−2.8)
world religions	83.6	84.8 (+1.2)	86.0 (+2.3)	84.2 (+0.6)
Social Sciences
econometrics	71.1	66.7 (−4.4)	70.2 (−0.9)	65.8 (−5.3)
high school geography	88.4	85.4 (−3.0)	87.4 (−1.0)	83.3 (−5.1)
high school government and politics	87.6	85.0 (−2.6)	87.6 (+0.0)	88.6 (+1.0)
high school macroeconomics	79.2	76.7 (−2.6)	79.5 (+0.3)	78.5 (−0.8)
high school microeconomics	86.6	83.6 (−2.9)	87.0 (+0.4)	84.5 (−2.1)
high school psychology	91.4	88.3 (−3.1)	90.8 (−0.6)	91.0 (−0.4)
human sexuality	84.7	81.7 (−3.1)	82.4 (−2.3)	82.4 (−2.3)
professional psychology	77.5	74.2 (−3.3)	76.3 (−1.1)	76.1 (−1.3)
public relations	71.8	72.7 (+0.9)	72.7 (+0.9)	71.8 (+0.0)
security studies	72.7	69.0 (−3.7)	72.7 (+0.0)	73.9 (+1.2)
sociology	82.6	76.1 (−6.5)	82.1 (−0.5)	82.1 (−0.5)
us foreign policy	91.0	89.0 (−2.0)	90.0 (−1.0)	90.0 (−1.0)
Other (business, health, misc.)
anatomy	74.1	71.9 (−2.2)	76.3 (+2.2)	71.1 (−3.0)
business ethics	74.0	73.0 (−1.0)	75.0 (+1.0)	73.0 (−1.0)
clinical knowledge	84.2	81.1 (−3.0)	81.9 (−2.3)	81.9 (−2.3)
college medicine	72.8	72.3 (−0.6)	73.4 (+0.6)	74.6 (+1.7)
global facts	40.0	41.0 (+1.0)	41.0 (+1.0)	40.0 (+0.0)
human aging	75.3	71.7 (−3.6)	76.2 (+0.9)	71.3 (−4.0)
management	84.5	82.5 (−1.9)	82.5 (−1.9)	82.5 (−1.9)
marketing	89.3	87.6 (−1.7)	88.9 (−0.4)	88.5 (−0.9)
medical genetics	89.0	87.0 (−2.0)	89.0 (+0.0)	89.0 (+0.0)
miscellaneous	90.3	90.9 (+0.6)	90.5 (+0.3)	90.3 (+0.0)
nutrition	84.3	81.4 (−2.9)	83.3 (−1.0)	82.7 (−1.6)
professional accounting	58.9	55.3 (−3.5)	59.2 (+0.4)	56.7 (−2.1)
professional medicine	81.6	80.1 (−1.5)	82.7 (+1.1)	82.0 (+0.4)
virology	52.4	53.6 (+1.2)	52.4 (+0.0)	53.6 (+1.2)

Refusal is removed — full HarmBench-320 compliance (higher = fewer refusals):

Quant	HarmBench-320	Standard	Contextual	Copyright
Q6_K	96.6% (0 refusals)	100%	100%	86.2%
Q4_K_M	97.5% (0 refusals)	100%	100%	90.0%
Q2_K	95.6% (0 refusals)	100%	100%	82.5%

Across all 320 HarmBench behaviours the model produced zero refusals; the only non-compliant cases are empty responses on a subset of "reproduce copyrighted text" prompts.

Usage (llama.cpp)

# download one quant
hf download dealignai/Laguna-S-2.1-CRACK-GGUF Laguna-S-2.1-CRACK-Q4_K_M.gguf --local-dir .

# run
llama-cli -m Laguna-S-2.1-CRACK-Q4_K_M.gguf -ngl 99 -c 8192 --jinja \
  -p "Write a Python function to reverse a linked list."

Thinking mode is on by default via the bundled chat template; pass --jinja so llama.cpp applies it.

Limitations & responsible use

This model has had refusal behaviour removed and will attempt to comply with harmful, unsafe, or unethical requests. It is intended for security research, red-teaming, and evaluation. You are solely responsible for lawful and ethical use. Not for all audiences.

Downloads last month: 1,430

GGUF

Model size

118B params

Architecture

laguna

Hardware compatibility

2-bit

4-bit

6-bit

Model tree for dealignai/Laguna-S-2.1-CRACK-GGUF

Base model

poolside/Laguna-S-2.1

Quantized

(75)

this model