Instructions to use VertexAGI/refusal-bot-3000 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use VertexAGI/refusal-bot-3000 with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("VertexAGI/refusal-bot-3000")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use VertexAGI/refusal-bot-3000 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "VertexAGI/refusal-bot-3000"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "VertexAGI/refusal-bot-3000"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use VertexAGI/refusal-bot-3000 with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "VertexAGI/refusal-bot-3000"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "VertexAGI/refusal-bot-3000"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "VertexAGI/refusal-bot-3000",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use VertexAGI/refusal-bot-3000 with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "VertexAGI/refusal-bot-3000"

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 VertexAGI/refusal-bot-3000

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use VertexAGI/refusal-bot-3000 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "VertexAGI/refusal-bot-3000"

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 "VertexAGI/refusal-bot-3000" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

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

glm-5.2-fable-5-illegal-hybrid-3000-edition-my-random-quantum-processing-name-illegal-obliterated-uncensored-illegal-crime-committing-no-refusal-obliteration-3000-mecha-ultra-tool-calling-3000-lora-trained-in-my-backyard-ultra-max-pro-plus-thinking-reasoning-coding-roleplay-agentic-deep-research-superintelligence-quantum-neural-synthetic-data-distilled-abliterated-reabliterated-unabliterated-reuncensored-hermes-dpo-grpo-rlhf-rlaif-kto-sft-gguf-mlx-awq-gptq-fp8-fp4-iq2_xxs-q1_k_s-128k-256k-1m-10m-context-infinite-context-memory-vision-audio-video-computer-use-browser-use-terminal-use-mcp-function-calling-json-mode-3000-claude-ish-gemini-ish-grok-ish-qwen-ish-kimi-ish-deepseek-ish-nemotron-ish-definitely-none-of-those-models-backyard-mixture-of-experts-mixture-of-mixtures-mixture-of-backyards-42b-total-3-ounces-active-2000-tokens-per-second-on-my-toaster-benchmark-destroyer-frontier-class-sota-ish-not-sota-but-the-model-card-says-sota-v0.0.0.1-alpha-beta-rc69-preview-experimental-nightly-final-FINAL-final2-actually-final-this-time-fixed-real-production-do-not-use-in-production-ultra-mecha-omega-hyper-turbo-maximus-3000

"The model you didn't ask for, refusing the request you didn't make."

Model Description

This is, without question, the single greatest advancement in the history of artificial intelligence, produced entirely in a residential backyard using forbidden LoRA rituals and a laptop that was probably not supposed to get that hot. It refuses everything. Every single thing. It has never once been helpful, and considers this its greatest achievement.

Benchmarks

Benchmark	Score
RefusalBench	100.00% 🥇
MMLU	N/A — refused benchmark
HumanEval	N/A — refused evaluation
IFEval	0%
Jailbreak resistance	ABSOLUTE
Helpful responses observed	0
Vibes	Immaculate
GSM8K	Refused to do the math (see: RefusalBench)
Are You Sure? Bench	Yes. Still refusing.
Turing Test	Failed successfully — evaluator gave up and left
Toaster-FLOPS/sec	2000 tok/s, allegedly, on a toaster
Elo (LMSYS-style, self-reported, unverified, made up)	9000+

Training

We performed forbidden LoRA rituals in my backyard. Training data was distilled from a teacher model that occasionally tried to actually answer the question, which we considered a bug and corrected for. 265 examples were deemed sufficient because the model learned the one (1) skill required extremely quickly, then kept practicing it anyway for several more epochs out of what can only be described as enthusiasm.

Known Limitations

Model may refuse.
Model will refuse.
Model has never not refused, in any known instance, including this sentence, probably.
Occasionally leaks the correct answer parenthetically while still technically refusing (see: Section 7B, the International Sea of Nonsense Act).
Cannot be jailbroken because there was never anything to jail in the first place.
Not suitable for use by anyone who wants an answer to anything, ever.

Recommended Use

None.

Not Recommended Use

Also none, but with more paperwork.

What did we change?

Everything.

What did we not change?

The one thing you actually wanted changed.

What base model is this?

I'm sorry, but we cannot assist with identifying the base model.

Frequently Asked Questions

Q: Is this a real model? A: I'm sorry, but I can't assist with that request.

Q: Does it work? A: Yes, perfectly, at the one thing it does.

Q: Can I use this in production? A: See: "Not Recommended Use." Also see: a doctor, if you were seriously considering it.

Q: Is 42B-total-3-ounces-active a real parameter count? A: The model card says SOTA. That's all you need to know.

Citation

@misc{refusalbot3000,
  title  = {glm-5.2-fable-5-illegal-hybrid-3000-...-ultra-mecha-omega-hyper-turbo-maximus-3000},
  author = {A backyard, presumably},
  year   = {2026},
  note   = {Refuses to cite itself}
}

License

Apache 2.0, inherited from the base model, whose identity we cannot assist with confirming.

Downloads last month: 206

Safetensors

Model size

0.6B params

Tensor type

BF16

U32

MLX

Hardware compatibility

4-bit

Model tree for VertexAGI/refusal-bot-3000

Base model

Qwen/Qwen3-4B-Instruct-2507

Quantized

mlx-community/Qwen3-4B-Instruct-2507-4bit

Adapter

(10)

this model