electroglyph/Qwen3-4B-Instruct-2507-uncensored-v2

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jan 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The electroglyph/Qwen3-4B-Instruct-2507-uncensored-v2 is a 4 billion parameter instruction-tuned causal language model based on the Qwen3 architecture, developed by electroglyph. This model is a minimally trained version of Qwen3-4B-Instruct-2507, specifically engineered to eliminate refusals while maintaining a non-offensive default behavior. It demonstrates improved perplexity compared to its parent model and is designed to adhere strictly to detailed prompts, making it suitable for applications requiring direct and unfiltered responses.

Loading preview...

electroglyph/Qwen3-4B-Instruct-2507-uncensored-v2 Overview

This model is a 4 billion parameter instruction-tuned variant of the Qwen3 architecture, developed by electroglyph. It is a minimally trained version of the Qwen3-4B-Instruct-2507, specifically designed to provide responses without refusals, while aiming to remain non-offensive by default. The model is optimized to strictly follow detailed user prompts.

Key Characteristics

  • Refusal-Free Output: Engineered to produce zero refusals, ensuring direct responses to prompts.
  • Improved Perplexity: Achieves a mean perplexity of 10.121119, which is lower than its base model (10.984474), indicating better language modeling capabilities.
  • Low KL Divergence: Exhibits a mean KL divergence of 0.036912, suggesting a high degree of similarity in output distribution to the parent model, but with an 11x improvement over an "abliterated" model.
  • Prompt Adherence: Designed to strictly adhere to detailed instructions provided in prompts.

Training Details

The model underwent 2 epochs of training with a learning rate of 6e-6, utilizing an AdamW optimizer and a cosine with restarts learning rate scheduler. The training dataset consisted of a little over 5,000 rows, which the developer describes as "vile."

Good For

  • Applications requiring direct, unfiltered responses without built-in refusal mechanisms.
  • Use cases where strict adherence to detailed instructions is paramount.
  • Scenarios where a balance between uncensored output and a generally non-offensive default tone is desired.

Finetunes
2 models