Model Overview

This model, megabytes/gemma-3-1b-qat-int4-heretic, is a 1 billion parameter instruction-tuned variant of Google's Gemma 3 multimodal language model. It is a decensored version of the original google/gemma-3-1b-it-qat-int4-unquantized model, created using the Heretic tool. A key differentiator is its significantly reduced refusal rate (6/100 compared to 88/100 for the original) while maintaining similar quality, achieved through Quantization Aware Training (QAT).

Key Capabilities

• Multimodal Input: Handles both text and image inputs, generating text outputs.
• Instruction-Tuned: Optimized for following instructions and generating coherent responses.
• Efficient Deployment: QAT enables similar quality to bfloat16 models with significantly reduced memory requirements, suitable for resource-constrained environments like laptops or desktops.
• Decensored: Modified to reduce content refusals compared to the base model.
• Context Window: Supports a 32K token context window for the 1B size.

Use Cases

• Content Creation: Generating creative text formats, marketing copy, or email drafts.
• Conversational AI: Powering chatbots and virtual assistants.
• Text Summarization: Creating concise summaries of documents.
• Image Data Extraction: Interpreting and summarizing visual data for text communications.
• Research & Education: Serving as a foundation for VLM and NLP research, or language learning tools.