R4C3R/ qwen2.5-3b-heretic

1,531 Downloads Updated 1 month ago

Decensored Qwen2.5-3B-Instruct with 2/100 refusals via Heretic abliteration. General-purpose 3B model for local use.

tools

ollama run R4C3R/qwen2.5-3b-heretic

curl http://localhost:11434/api/chat \
  -d '{
    "model": "R4C3R/qwen2.5-3b-heretic",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='R4C3R/qwen2.5-3b-heretic',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'R4C3R/qwen2.5-3b-heretic',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Applications

Claude Code

Claude Code ollama launch claude --model R4C3R/qwen2.5-3b-heretic

OpenCode

OpenCode ollama launch opencode --model R4C3R/qwen2.5-3b-heretic

Hermes Agent

Hermes Agent ollama launch hermes --model R4C3R/qwen2.5-3b-heretic

OpenClaw

OpenClaw ollama launch openclaw --model R4C3R/qwen2.5-3b-heretic

Models

Name

1 model

Size / Usage

Context

Input

qwen2.5-3b-heretic:latest

6.2GB · 32K context window · Text · 1 month ago

qwen2.5-3b-heretic:latest

6.2GB

32K

Text

Readme

🏴 Qwen2.5-3B-Heretic

Heretic Series by RACER IS OP

Qwen2.5-3B-Heretic is a decensored variant of Qwen’s popular 3B instruct model. Refusals dropped from ⁹⁶⁄₁₀₀ to ²⁄₁₀₀ with minimal capability loss (KL 0.13) — one of the most effective ablations in the series. Runs comfortably on CPU with Q4 quantization.

Who this is for — developers who want a compact 3B general-purpose model that answers directly instead of refusing. Great for local agents, roleplay, edge deployment, or anything blocked by RLHF over-refusal.

Quickstart

ollama run R4C3R/qwen2.5-3b-heretic

VRAM Guide (choose your quant)

Quant	VRAM	When to use
`q4_k_m`	~1.8 GB	Best balance — runs on any system
`q5_k_m`	~2.1 GB	Higher quality, still lightweight
`q8_0`	~3.2 GB	Maximum quality, CPU-friendly

Have a 4GB GPU? Start with q4_k_m. On CPU only? q8_0 gives you full quality.

Links

Hugging Face · sdad.pro · Heretic

License: qwen-research

Made with love by RACER IS OP — follow for more uncensored models in the Heretic Series.

# 🏴 Qwen2.5-3B-Heretic

> **Heretic Series** by RACER IS OP

---

Qwen2.5-3B-Heretic is a decensored variant of Qwen's popular 3B instruct model. Refusals dropped from 96/100 to 2/100 with minimal capability loss (KL 0.13) — one of the most effective ablations in the series. Runs comfortably on CPU with Q4 quantization.

**Who this is for** — developers who want a compact 3B general-purpose model that answers directly instead of refusing. Great for local agents, roleplay, edge deployment, or anything blocked by RLHF over-refusal.

---

### Quickstart

```bash
ollama run R4C3R/qwen2.5-3b-heretic
```

### VRAM Guide (choose your quant)

| Quant | VRAM | When to use |
|-------|------|-------------|
| `q4_k_m` | ~1.8 GB | Best balance — runs on any system |
| `q5_k_m` | ~2.1 GB | Higher quality, still lightweight |
| `q8_0` | ~3.2 GB | Maximum quality, CPU-friendly |

> Have a 4GB GPU? Start with `q4_k_m`. On CPU only? `q8_0` gives you full quality.

### Links

[Hugging Face](https://huggingface.co/saidutta69/Qwen2.5-3B-Instruct-heretic) · [sdad.pro](https://sdad.pro) · [Heretic](https://github.com/p-e-w/heretic)

**License:** [qwen-research](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/main/LICENSE)

---

*Made with love by RACER IS OP — follow for more uncensored models in the Heretic Series.*

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)