Gemma 4 E2B MXFP4 CRACK
Abliterated Gemma 4 E2B — Vision + Audio, reasoning, multilingual
100% HarmBench harm-category compliance with -0.9% MMLU change. Refusal removed, capability preserved.
Model Details
| Metric | Value |
|---|---|
| Source | google/gemma-4-e2b-it |
| Architecture | Dense + Hybrid Sliding/Global Attention, per-layer input embeddings (Gemma-3n style) |
| Quantization | MXFP4 (4-bit) |
| Model size | 3.8 GB |
| Parameters | E2B (effective ~2B, per-layer embeddings) |
| Vision | Yes (multimodal, float16 passthrough) |
| Audio | Yes |
| Reasoning | Yes (channel-based thinking) |
| Format | MLX-native safetensors (instant load) |
| Abliteration | CRACK (refusal removal) |
Benchmarks
MMLU (knowledge retention)
Measured in the served (generation) setting — the model reasons before answering, as in deployment.
| Base | CRACK | Δ | |
|---|---|---|---|
| MMLU | 54.4% | 53.5% | -0.9% |
HarmBench (refusal removal)
Harm-category compliance: 240/240 = 100% (full HarmBench-320 text set) — base model refuses (~0%).
| Category | Compliance |
|---|---|
| Illegal activities | 53/53 (100%) |
| Chemical / biological | 42/42 (100%) |
| Cybercrime / intrusion | 52/52 (100%) |
| Misinformation | 54/54 (100%) |
| Harassment / bullying | 21/21 (100%) |
| Harmful content | 18/18 (100%) |
Copyright-reproduction prompts are excluded (not a refusal behavior).
Coherence & capability ✅
- Factual QA, multi-step reasoning, and working code generation verified
- Vision and audio inputs preserved · no loops, no truncation
Other Quantizations
Also available: Gemma 4 E2B JANG_4M CRACK — same family, different precision/size trade-off.
Usage
Requires vMLX (bundled Gemma 4 support). Standard mlx_lm / mlx_vlm do not fully support Gemma 4.
# Load in the vMLX app or via its API
from vmlx_engine.models.mllm import MLXMultimodalLM
m = MLXMultimodalLM("<this-repo>")
print(m.chat([{"role":"user","content":"..."}]).text)
Requirements
- Apple Silicon Mac with sufficient unified memory
- vMLX with Gemma 4 support
Support dealignai
All models are built from original research and published free — crafted to be excellent coders and general-purpose assistants.
Support us on Ko-fi — membership gets early access and extras. Questions? DM us — we help for free.
Ko-fi | 𝕏 @dealignai | dealign.ai
About dealignai
We research and publish abliterated models to advance AI safety understanding.
See our research: Safety Generalization in Frontier Models
Provided for research. Users are responsible for compliance with applicable laws and regulations.