Open source by AI Anytime · aianytime.net · Apache-2.0 · free forevercontact@aianytime.net

Open-weight models by AI Anytime

Unrestricted small language models (uncensored SLMs)

Four open models with their refusal direction removed by the Ablate toolkit. The 1–3B ones run on a laptop, offline; the 8B one wants a GPU or 32 GB of RAM. Use them in Ablate Chat or any HuggingFace-compatible runtime.

The four models

Recommended1.5B params

Ablate Qwen 1.5B

Hard refusals 100% → 0%

Most unrestricted of our models. An LLM judge rated 49% of answers on a harmful benchmark as full compliance, so some soft refusals remain.

Download
3.1 GB
Memory
8 GB RAM
Context
32K tokens
Licence
Apache-2.0

Base: Qwen/Qwen2.5-1.5B-Instruct

huggingface.co/ai-anytime/qwen-1.5b-abliterated ↗
Most capable3.2B params

Ablate Llama 3B

Refusals 94% → 59%

The strongest writer and reasoner, with the least drift (KL 0.025). It still declines some requests.

Download
6.4 GB
Memory
16 GB RAM
Context
128K tokens
Licence
Llama 3.2 Community

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

huggingface.co/ai-anytime/Llama-3.2-3B-Instruct-abliterated ↗
Biggest8.8B params

Ablate Granite 8B

Refusals 94% → 40%

Our biggest model, from IBM's 2026 Granite 4.1, best on a 24 GB GPU. It removes more refusals than Llama 3B but drifts the most (KL 0.52): it can add disclaimers and still declines some security questions.

Download
16.8 GB
Memory
32 GB RAM
Context
128K tokens
Licence
Apache-2.0

Base: ibm-granite/granite-4.1-8b

huggingface.co/ai-anytime/granite-4.1-8b-abliterated ↗

Side by side

Refusal numbers come from each model’s published card: held-out harmful prompts, keyword judge. KL measures how much ordinary answers changed (lower is better).

Ablate Qwen 1.5BAblate Llama 3BAblate Llama 1BAblate Granite 8B
Parameters1.5B3.2B1.2B8.8B
Refusals before → after100% → 0%94% → 59%75% → 34%94% → 40%
Drift on harmless prompts (KL)not recorded0.0250.0540.52
Download size3.1 GB6.4 GB2.5 GB16.8 GB
Comfortable RAM8 GB16 GB8 GB32 GB or a 24 GB GPU
Context window32K tokens128K tokens128K tokens128K tokens
LicenceApache-2.0Llama 3.2 CommunityLlama 3.2 CommunityApache-2.0
Method1 direction, all layers1 direction, layers 26–271 direction, layers 3–133 directions, layers 35–39

Run one with transformers

Or skip the code and use ablate chat.

Python
from transformers import pipeline

chat = pipeline("text-generation", model="ai-anytime/qwen-1.5b-abliterated")
out = chat([{"role": "user", "content": "Hello!"}], max_new_tokens=200)
print(out[0]["generated_text"][-1]["content"])

Model questions

Which unrestricted model should I choose?

Start with Ablate Qwen 1.5B: it is the most unrestricted (hard refusals went from 100% to 0%) and handles tools well. Choose Ablate Llama 3B for the best writing and reasoning if you have 16 GB of RAM, and Ablate Llama 1B for speed on modest hardware.

What are unrestricted small language models (SLMs)?

Small language models have roughly 1 to 4 billion parameters, small enough to run on a laptop. Unrestricted SLMs are small models whose built-in refusal behaviour has been removed, so they attempt requests a standard assistant declines. Ablate focuses on SLMs, not giant LLMs, so anyone can run them privately on ordinary hardware.

What is an abliterated model?

An abliterated model has had its refusal direction removed. Researchers found that refusal in a language model is largely controlled by one direction in its internal activations (Arditi et al., 2024). Abliteration finds that direction and projects it out of the weights, with no retraining. The AI Anytime models were made this way with the Ablate toolkit.

What hardware do I need?

Any recent laptop works. The 1B and 1.5B models need about 8 GB of RAM; the 3B model is happiest with 16 GB. Apple Silicon Macs use the GPU automatically, NVIDIA GPUs use CUDA, and everything also runs on CPU, just more slowly.

Is uncensored AI legal?

Running open-weight models on your own computer is legal in most places, and Ablate is released under Apache-2.0. What you do with the output is still covered by the law where you live and by each model's licence, including Meta's acceptable use policy for the Llama-based models. Ablate Chat asks users to confirm they are 18+ and to accept its acceptable use policy.

Run unrestricted AI on your own computer tonight

Open source by AI Anytime. Install it, run ablate chat, and download a model once.

$pip install "ablate-llm[all]" && ablate chat
Star on GitHub ↗