pmarreck

pmarreck / gemma4-heretical

Public

Abliterated Gemma 4 31B for Ollama and MLX — one-command setup with correct chat template fix

85
6
89% credibility
Found Apr 08, 2026 at 85 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Shell
AI Summary

This repository offers an easy one-command setup to run a modified, uncensored version of Google's Gemma 4 31B AI model locally on computers using Ollama or natively on Apple Silicon.

How It Works

1
📰 Discover Uncensored AI

You hear about a powerful, free AI model that answers any question without refusing, unlike others.

2
💾 Grab the Setup

Download the simple package that makes installing this AI super easy.

3
Pick Your Computer Type
🖥️
Regular Computer

Set it up to run smoothly on most computers.

🍎
Apple Mac

Get a super-fast version just for Apple computers.

4
🚀 Install with One Click

Run the easy setup and watch your personal AI come to life on your machine.

5
⚖️ Choose the Right Size

Pick a lighter version if your computer has less memory, so it runs perfectly.

6
💬 Start Chatting Freely

Open the AI and ask anything – it responds openly without holding back.

🎉 Your AI is Ready!

Enjoy your uncensored assistant that thinks and talks like you want, all local on your computer.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 85 to 85 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is gemma4-heretical?

This Shell script repo delivers a one-command setup for running an abliterated Gemma 4 31B model—uncensored via Heretic's Arbitrary-Rank Ablation—in Ollama or MLX on Apple Silicon. It pulls community GGUF weights, fixes the broken chat template that plagues stock Gemma uploads (causing endless "---" outputs), and registers everything correctly for seamless chatting. Developers get a refusal rate slashed from 98 to 5 out of 100, with near-zero quality loss via KL divergence of 0.012.

Why is it gaining traction?

Unlike raw abliterated Gemma 3 12B/27B or 4B ports on Ollama, this handles Gemma 4 31B's specific renderer/parser quirks out of the box—no manual tweaks needed. The hook is instant quant options like Q4_K_M (~19GB RAM) or IQ4_NL, plus MLX conversion for native Metal inference and OpenAI-compatible APIs. It's a quick win over fiddly deepseek-r1-abliterated github setups or gemma-obliterated GGUF hunts.

Who should use this?

Ollama power users experimenting with abliterated LLMs for uncensored roleplay or research prompts. Apple Silicon devs needing fast local inference without cloud refusals, like prototyping chat apps. AI tinkerers short on RAM who want Gemma 3-abliterated vibes but scaled to 31B.

Verdict

Grab it if you're chasing low-refusal Gemma 4 locally—solid docs and 85 stars show promise, though the 0.8999999761581421% credibility score flags early maturity; test quantizations first. Skip for production until more battle-testing.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.