Instructions to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

Use Docker

docker model run hf.co/Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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": "Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

Ollama
How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with Ollama:
```
ollama run hf.co/Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M
```
Unsloth Desktop

How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with Docker Model Runner:
```
docker model run hf.co/Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M
```

Lemonade

How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Llama-3.2-3B-RabbitSoftware-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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 "Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-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"

Llama-3.2-3B-RabbitSoftware (GGUF, Q4_K_M)

The model behind RabbitSoftware.inc, the assistant of the RabbitSoftware OS (DNA-Blockchain/Helloworld). The OS runs a network of research nodes that collect public biomedical records into a local corpus and record them on a signed, replicated chain. This model is the reasoning layer on top of that corpus. It produces cited technical answers from the retrieved records, and it reports on the OS's own chain, swarm and integrity results.

Model


Base model	Meta Llama 3.2 3B Instruct
Fine-tuning	LoRA, round 2, trained by distillation: Qwen2.5-7B-Instruct generated candidate explanations, which were filtered by automatic fact checks before training the 3B model.
Export	LoRA merged into the base weights and quantized with llama.cpp to GGUF Q4_K_M (2.0 GB).
File	`nos-lora.Q4_K_M.gguf`, SHA-256 `8aa05e2879f4e1e37395a238497726c2454fc0dc66661f989996e1b0dca473ef`

Pipeline it runs in

Ingestion: research nodes query PubMed, Europe PMC, NIH RePORTER and ClinicalTrials.gov. Records are saved to the local catalog, and each batch is published to the shared research chain as a signed public_research_records event.
Retrieval: a question is matched against the catalog by keyword (catalog score ≥ 3) and by meaning, using nomic-embed-text embeddings with cosine ≥ 0.62, or TF-IDF with cosine ≥ 0.12 when no embedding model is installed. Duplicates across sources are merged, and the top 5 records go to the model.
Generation: the model receives each record's title, source, year and up to 900 characters of abstract. It answers in three sections:
- Findings: the reported figures, quoted exactly;
- Methods and evidence: each cited record's study type and the strength of evidence;
- Limitations: what the records don't establish, and where they disagree.
Every claim is cited [n]. Citations to records that weren't supplied are removed after generation.
Reporting: the answer is returned with its sources and a retrieval line covering match methods, similarity range and cutoff, sources, years and abstract coverage.

Where its answers come from

Besides the public research corpus, the assistant answers from the project's own knowledge base: the source-linked research reports in docs/research/, split into sections and indexed by meaning. Those answers carry [K1]-style citations naming the report section and its file, so a claim can be traced to the report and from there to its primary source. The reports cover EEG decoding and DNA editing in cancer; they are research summaries, not medical advice.

Download and use

Each version is recorded in GitHub, in deploy/huggingface/model-versions.json: its SHA-256, the commit it was built from, its training and its scores. Released versions are tagged model-v<version> here. New ones wait on the candidates branch until the RabbitSoftware release that ships them.

With RabbitSoftware (downloads with your Hugging Face login, checks the SHA-256, adds it to Ollama):

hf auth login
ollama pull llama3.2:3b
python rabbit.py model install            # or --version 1.0.0
ollama run rabbitsoftware

With Ollama alone: a private repo needs Ollama's public key (~/.ollama/id_ed25519.pub) added under Settings → SSH Keys on Hugging Face. Use the full host name; hf.co/ fails with a "realm host" error on private repos:

ollama run huggingface.co/<your-hf-username>/Llama-3.2-3B-RabbitSoftware-GGUF:Q4_K_M

With llama.cpp:

hf download <your-hf-username>/Llama-3.2-3B-RabbitSoftware-GGUF nos-lora.Q4_K_M.gguf --revision model-v1.0.0 --local-dir .
llama-cli -m nos-lora.Q4_K_M.gguf -cnv

Use Llama 3.2's chat format (the GGUF carries it). The model was tuned to write from supplied facts and records, so give it the facts in the prompt rather than asking it to recall them.

Serving

Server: a Hugging Face Inference Endpoint (rabbitsoftware-model) running the llama.cpp server on one NVIDIA T4. It scales to zero after 15 minutes idle.
API: OpenAI-compatible POST /v1/chat/completions, model name rabbitsoftware. Answers are capped at 400 tokens.
Access: the endpoint is private. The owner's installation authenticates with its own Hugging Face login. Other installations go through a gateway on Cloudflare Workers, which holds the token and enforces per-user and daily limits.
Consent: RabbitSoftware.inc asks before every question it sends to the server. Declining answers with the model on the local machine.
Throughput: about 84 tokens/s when warm; 30–60 s cold start after scale-to-zero.

Evaluation and limits

Measured on this exact file (SHA-256 8aa05e28…) on 2 October 2026 with python local_ai_tuning.py run --variants system,lora, at temperature 0:

Task	This model	Same base model without the fine-tune	Most common problems
Explaining computed DNA-analysis facts	6/8	6/8	2 answers contradicted the facts
Summarizing a research title	2/12	8/12	8 copied the title, 2 dropped names

On explanations it matches the prompt-engineered model. Round 2 improved exact preservation of numbers. On summaries the fine-tune is worse than no fine-tune, so RabbitSoftware writes summaries with the prompt-engineered nos-summary (no fine-tune) and uses this model for explanations only.

It's a 3B model and can be wrong even when it cites a record. Known errors from testing:
- it reversed the direction of the sickle-cell substitution (the correct direction is glutamic acid → valine at β-globin position 6);
- it over-read a survey paper as a treatment approval.
It answers only from the records supplied. When nothing relevant is retrieved, the assistant offers a search instead of answering.
Not medical advice. Every answer in RabbitSoftware.inc says so.

Data

Training used public research text and synthetic examples, with no personal data. Later rounds may add answers that users explicitly chose to share. These are screened for personal information and carry no name, account or device identifier (see the private dataset rabbitsoftware-training).

License

Downloads last month: 3

GGUF

Model size

3B params

Architecture

llama

Hardware compatibility

4-bit

Model tree for Therealsickonechase-bit/Llama-3.2-3B-RabbitSoftware-GGUF

Base model

meta-llama/Llama-3.2-3B-Instruct

Adapter

(851)

this model