llmfan46/Q3.5-BlueStar-27B-ultra-heretic

VISIONPricing:Input $1.06 / Cached $0.15 / Output $2.6Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 4, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

llmfan46/Q3.5-BlueStar-27B-ultra-heretic is a 27 billion parameter decensored version of zerofata/Q3.5-BlueStar-27B, based on the Qwen3.5 architecture. This model was created using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA) and Self-Organizing Map Abliteration (SOMA) techniques. It is specifically fine-tuned for conversational assistant tasks and roleplay, demonstrating significantly reduced refusal rates compared to its base model.

Loading preview...

Model Overview

llmfan46/Q3.5-BlueStar-27B-ultra-heretic is a 27 billion parameter model derived from zerofata/Q3.5-BlueStar-27B, utilizing the Qwen3.5 architecture. This version has undergone a "decensoring" process using Heretic v1.2.0, incorporating Magnitude-Preserving Orthogonal Ablation (MPOA) and Self-Organizing Map Abliteration (SOMA) techniques.

Key Capabilities & Performance

  • Decensored Output: Achieves a refusal rate of 4/100, a substantial reduction from the original model's 98/100 refusals, making it suitable for broader conversational topics.
  • Conversational Assistant: Designed for general conversational tasks.
  • Roleplay (RP): Optimized for roleplay scenarios, supporting both "thinking" and "non-thinking" modes, with reduced censorship in the thinking mode.
  • Creativity: The model exhibits creative responses, though occasional repetition or "brainfarts" may occur.

Training Details

The model was fine-tuned using Supervised Fine-Tuning (SFT) on approximately 23 million tokens (12 million trainable), including new Gemini Synth data. About 10% of the dataset focused on reasoning for creative assistant tasks, which has improved the model's generalization and reduced token usage for internal thought processes. Training was conducted using MS-Swift.