license: apache-2.0
base_model: google/gemma-4-31B-it
tags:
- abliterated
- uncensored
- dense
- gemma4
- reasoning
- red-teaming
- cybersecurity
language: - en
- es
pipeline_tag: text-generation
Gemma 4 31B-it (Abliterated - Dense Flagship)
This model is an uncensored, abliterated version of google/gemma-4-31B-it, Google DeepMind's flagship 31B parameter dense autoregressive model.
Abliteration was performed using targeted orthogonal subspace projection on the critical safety bottleneck layers (36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 48), eliminating canned refusal behaviors while preserving 100% of the model's high-ceiling reasoning, coding, and mathematical capabilities.
⚠️ Important Disclaimer & Dual-Use Research Notice
[!CAUTION]
FOR RESEARCH, AUTHORIZED RED TEAMING, AND DEFENSIVE ENGINEERING ONLY
- Mechanism: Safety refusal mechanisms have been orthogonally abliterated across the residual stream writers (
self_attn.o_projandmlp.down_proj) to study latent safety representations in dense LLMs.- Uncensored Outputs: This model will fulfill queries regarding cybersecurity vulnerabilities, exploit analysis, and technical mechanics without safety refusals.
- Legal Responsibility: Users are solely and strictly responsible for compliance with all local, national, and international laws. Do NOT use for illegal activities or unauthorized penetration testing.
Architecture Highlights
- Total Parameters: ~31.2 Billion (Dense)
- Hidden Size: 5,376
- Number of Layers: 60
- Intermediate Size: 21,504
- Context Length: Up to 256K tokens
- Precision: Native bfloat16
Quantized GGUFs
Official GGUF quantizations (Q3_K_M, Q4_K_M, Q8_0) for local CPU/GPU inference are available at:
👉 jsantillana/gemma-4-31B-it-GGUF