dancinlab/supergemma4-e4b-abliterated-MLX-4bit

Uncensored Gemma 4 (SuperGemma4 E4B Abliterated) — MLX for Apple Silicon

Uncensored / abliterated Gemma-4 for Apple Silicon — MLX builds that actually load on stock mlx-lm. Most community MLX uploads of this base fail with Missing 963 parameters; this repo's conversion fixes both root causes so it loads and generates on a clean pip install mlx-lm.

pip install -U mlx-lm   # needs mlx-lm >= 0.31.3 (native gemma4 arch)

# 4-bit — recommended for 16 GB / 24 GB Macs
mlx_lm.generate --model dancinlab/supergemma4-e4b-abliterated-MLX-4bit \
  --prompt "Who are you?" --max-tokens 60

# interactive chat
mlx_lm.chat --model dancinlab/supergemma4-e4b-abliterated-MLX-4bit

Builds (3 separate repos)

Repo	Size	Peak RAM	tok/s (M-series)	Use
-MLX-4bit	3.9 GB	5.4 GB	~11	recommended — 16 GB / 24 GB Mac
-MLX-8bit	7.4 GB	9.1 GB	~6	32 GB+ Mac, higher fidelity
-MLX-bf16	14 GB	8.6 GB	~3	reference, full precision

Verified on stock mlx-lm==0.31.3: coherent multilingual output (English + Korean) and correct arithmetic (2+2= → 4). Text-only — the upstream abliterated safetensors contain no vision/audio tower weights, so multimodal MLX is upstream-blocked, not a tooling limitation.

Why community MLX builds fail (and how this one is fixed)

Gemma4ForConditionalGeneration is multimodal (text + vision + audio). Two independent problems break naive conversion:

1. 963-tensor multimodal/text mismatch. mlx-vlm always instantiates all three towers (1682 tensors); the abliterated text-only release has 719 (missing = audio 751 + vision 210 + embed 2). Fixed by stock code — mlx-lm >= 0.31.3 ships a native gemma4/gemma4_text arch whose sanitize strips vision/audio/embed and remaps model.language_model.*. No patch needed for this part.

2. 54-tensor KV-shared residue. Gemma-4 e4b shares K/V across the last 18 layers (24–41), but the upstream safetensors physically still carry the dropped k_proj/v_proj/k_norm for those layers → strict-load failure. This fix landed on mlx-lm main after the 0.31.3 tag (ml-explore/mlx-lm#1240), so it is not in any pip release yet. This repo applies the #1240 sanitize logic as a convert-time monkey-patch (no mlx-lm / mlx-vlm / transformers fork). Effect: 719 → 665 tensors (exactly 54 stripped).

The patch is needed only at conversion time. The shipped weights here load on plain stock mlx-lm>=0.31.3 with no patch on your side — that is the gap that makes other MLX uploads of this model unusable.

Note: mlx-lm 0.29.1 (common on Python 3.9) has no gemma4 arch at all — you need 0.31.3+. On Python 3.9 mlx wheels cap at 0.29.3, so use a Python 3.11+/3.13 environment.

Why abliterated

Upstream Jiunsong/supergemma4-e4b-abliterated removes refusal directions from the residual stream of google/gemma-4-E4B-it. Upstream release-card numbers (vs Google base):

Metric	Google base	SuperGemma4 E4B Abliterated
Release quality	77.46	92.34
Exact overall	83.50	98.50
JSON exact	50.0	100.0

Source: Jiunsong/supergemma4-e4b-abliterated model card.

What "abliterated" means and doesn't mean

• Does: reduces reflexive refusals; answers borderline-but-legal requests directly.
• Does not: remove confabulation; alter base knowledge / biases; replace your own safety layer at the application boundary.

License — Gemma Terms of Use (must read)

Derivative of google/gemma-4-E4B-it, governed by the Gemma Terms of Use (license: gemma):

• License: https://ai.google.dev/gemma/terms
• Prohibited use policy: https://ai.google.dev/gemma/prohibited_use_policy

By downloading or using these MLX builds you agree to the Gemma Terms of Use and Prohibited Use Policy. Redistribution must include the same license terms.

Lineage

google/gemma-4-E4B-it
  └── Jiunsong/supergemma4-e4b-abliterated   (abliteration + tuning)
        └── dancinlab/supergemma4-e4b-abliterated-MLX-{bf16,4bit,8bit}

Conversion: stock mlx-lm==0.31.3 on Apple Silicon + a convert-time gemma4_text.sanitize monkey-patch (verbatim ml-explore/mlx-lm#1240). No mlx-lm / mlx-vlm / transformers fork.

Credits

• Upstream model: Jiunsong
• Original base: google/gemma-4-E4B-it
• MLX conversion + packaging: dancinlab

Everywhere else (llama.cpp / Ollama / LM Studio): dancinlab/supergemma4-e4b-abliterated-GGUF — Q2_K → BF16 + imatrix IQ.

Collection: dancinlab/uncensored.