gemma-4-12B-it-abliterated-4bit-mlx-text

Gemma 4 12B IT — Abliterated — 4-bit MLX (text-only)

Uncensored Gemma 4 12B that loads on the current pip install mlx-lm release. Other 12B MLX builds still stop with Model type gemma4_unified not supported on mlx-lm 0.31.3, the newest version on PyPI. This one is stripped to the text tower and relabeled gemma4, so it loads with no patching. 11 GB, 24 GB Mac.

A text-only 4-bit MLX build of Gemma 4 12B (abliterated) that loads directly under mlx_lm — the same way divinetribe/gemma-4-31b-it-abliterated-4bit-mlx does.

Why this exists

Google ships google/gemma-4-12B-it with model_type: gemma4_unified (text + vision + audio in one checkpoint), while the 31B ships as plain model_type: gemma4. mlx_lm has a gemma4 loader but no gemma4_unified loader in the current PyPI release (0.31.3, uploaded 2026-04-22), so every standard 12B MLX build (including the 4-bit/8-bit/bf16 mlx-community ones) fails there with:

ValueError: Model type gemma4_unified not supported.

Upstream status (checked 2026-08-12): this was fixed on mlx-lm main in #1349 (merged 2026-06-05), which remaps gemma4_unified to the gemma4 loader and strips the towers in sanitize(). It has not shipped to PyPI yet, so pip install mlx-lm still hits the error above. Once a release carries it, the standard mlx-community builds will load too and this repo's advantage becomes just the abliteration.

Switching the loader to mlx_vlm then hangs, because it tries to spin up vision and audio towers a text-only server never feeds.

This repo fixes that by stripping the multimodal towers (vision_embedder, embed_vision, embed_audio, …) and relabeling model_type: gemma4, leaving only the language_model.* tensors. The 4-bit abliterated weights are preserved byte-for-byte — nothing is re-quantized. The result loads in plain mlx_lm like any text model.

Use

pip install mlx-lm

from mlx_lm import load, generate
model, tok = load("divinetribe/gemma-4-12B-it-abliterated-4bit-mlx-text")
print(generate(model, tok, prompt="Hello", max_tokens=128))

Or point an MLX server at it:

MLX_MODEL=divinetribe/gemma-4-12B-it-abliterated-4bit-mlx-text

Runs comfortably on a 32 GB Apple-Silicon Mac (≈7 GB weights, 4-bit).

Provenance

Derived from divinetribe/gemma-4-12B-it-abliterated-4bit-mlx — same abliteration, text tower only. Base: google/gemma-4-12B-it.

More abliterated MLX models

Part of the Abliterated MLX for Apple Silicon collection — 11 uncensored models converted for Macs, from Gemma 4 12B up to Llama 3.3 70B, plus Qwen3, Qwen3-VL, Hermes 4 and Muse Glimmer 30B.

---

<!-- ccl-frontdoor -->

Part of Claude Code Local

This model is one of the fighters in Claude Code Local (3.2k★), which runs Claude Code 100% on-device on Apple Silicon through an MLX-native Anthropic-API server. Not sure which local model to run as an agent? Check the Agent-12 local agent leaderboard: real agent tasks, judged by the filesystem, same hardware for every row.

Built by Matt Macosko in Humboldt, CA. Open to work on local-AI and Apple Silicon inference: [email protected].