MobiusDevelopment/gemma4-E2B-it-qat-Q4-unsloth-heretic

VISIONConcurrency Cost:1Model Size:5.1BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 23, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

MobiusDevelopment's gemma4-E2B-it-qat-Q4-unsloth-heretic is a 5.1 billion parameter instruction-tuned multimodal language model based on Google DeepMind's Gemma 4 E2B architecture, featuring a 128K token context length. This model has been 'decensored' using the Heretic tool, removing refusal behaviors while retaining its core capabilities. It excels at multimodal tasks including text, image, and audio understanding, and is optimized for efficient deployment with its effective 2.3B parameter count.

Loading preview...

Model Overview

This model, gemma4-E2B-it-qat-Q4-unsloth-heretic, is a decensored version of Google DeepMind's Gemma 4 E2B instruction-tuned model. It leverages Unsloth's optimized base weights and has undergone "abliteration" using the Heretic tool to remove built-in safety refusals, allowing it to answer prompts it would otherwise decline. The model maintains its general capabilities, including multimodal understanding (text, image, audio) and a large context window.

Key Features & Capabilities

  • Decensored Behavior: Refusal behaviors have been removed, enabling the model to respond to a wider range of prompts.
  • Multimodal: Supports text, image, and audio input, generating text output. Audio capabilities are present despite some badge limitations on platforms like Ollama.
  • Efficient Architecture: Based on Gemma 4 E2B, designed for efficient on-device and edge deployment with an effective 2.3B parameters (5.1B total).
  • Long Context: Features a 128K token context length, utilizing a hybrid attention mechanism with Proportional RoPE (p-RoPE).
  • Core Capabilities: Includes built-in step-by-step reasoning, image understanding (OCR, chart analysis), speech recognition, function calling, and coding assistance.
  • Multilingual Support: Pre-trained on 140+ languages.

Abliteration Details

The decensoring process used Heretic v1.4.0 with Arbitrary-Rank Ablation (ARA) and row-norm preservation (LoRA rank 3). This method targets language-model output projections (o_proj, down_proj) to remove refusal directions while minimizing impact on other model functions. The selected abliteration trial resulted in 11/100 refusals and a KL divergence of 0.0489, indicating effective refusal removal with minimal output distribution drift.

Important Considerations

Users are responsible for the content generated, as the model's safety refusals have been removed. It will attempt to comply with prompts, including potentially harmful ones, that the original model would have refused. Deployment requires independent content-safety safeguards.

Finetunes
1 models