ressl/gemma-4-31B-it-uncensored

VISIONConcurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The ressl/gemma-4-31B-it-uncensored model is an uncensored build of Google's Gemma-4-31B-it instruction-tuned multimodal language model. Developed by Robert Ressl, this 30.7 billion parameter model achieves 0/686 effective refusals on harmful prompts across four independent datasets, while preserving the base model's capabilities with a low KL divergence of 0.1234. It is specifically designed for use cases requiring an assistant that does not refuse requests, such as ethical hacking, security research, and penetration testing.

Loading preview...

Overview

ressl/gemma-4-31B-it-uncensored is a specialized build of Google's gemma-4-31B-it model, developed by Robert Ressl. This 30.7 billion parameter multimodal model has been modified to be genuinely uncensored, demonstrating 0/686 effective refusals on harmful prompts across four independent datasets (JailbreakBench, tulu-harmbench, NousResearch/RefusalDataset, mlabonne/harmful_behaviors). This contrasts sharply with the base model's 99/100 refusal rate.

Key Capabilities & Features

  • Uncensored Responses: Provides compliant answers to requests that stock models typically refuse, with a 0% effective refusal rate.
  • Capability Preservation: Maintains the base model's coherence, mathematical, multilingual, and factual answering abilities, indicated by a low KL divergence of 0.1234.
  • Multimodal Support: Inherits multimodal capabilities from the base Gemma-4-31B-it model.
  • Broad Compatibility: Validated across various inference engines including vLLM, SGLang, and llama.cpp, and available in multiple formats (BF16, NVFP4, GGUF, MLX).
  • Weight-Level Uncensoring: The uncensoring is a fundamental weight-level change, ensuring it persists across different quantizations (e.g., 4-bit NVFP4, 2-bit GGUF).

Intended Use Cases

This model is particularly useful for applications where an unrefusing assistant is critical, such as:

  • Ethical Hacking & Security Research: Analyzing malware, exploit code, writing detection rules, and studying attack techniques.
  • Penetration Testing: Red-teaming exercises and vulnerability review without model refusal.
  • Research & Development: Scenarios requiring unrestricted model output for analysis or experimentation.

Finetunes
8 models