Bobi099/Qwen3.5-27B-heretic

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Feb 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Bobi099/Qwen3.5-27B-heretic is a 27 billion parameter multimodal causal language model, a decensored version of Qwen/Qwen3.5-27B created using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA). It features a unified vision-language foundation, an efficient hybrid architecture, and scalable reinforcement learning generalization, excelling in multimodal reasoning, coding, and agent tasks with a native context length of 262,144 tokens.

Loading preview...

Overview

Bobi099/Qwen3.5-27B-heretic is a 27 billion parameter multimodal causal language model, derived from Qwen/Qwen3.5-27B through a decensoring process using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA). This modification significantly reduces refusals from 94/100 in the original model to 14/100, while maintaining a low KL divergence of 0.0653. The base Qwen3.5 model integrates breakthroughs in multimodal learning, architectural efficiency, and reinforcement learning, offering a native context length of 262,144 tokens, extensible up to 1,010,000 tokens via YaRN scaling.

Key Capabilities

  • Decensored Responses: Achieves a substantial reduction in refusal rates compared to the original Qwen3.5-27B model.
  • Unified Vision-Language Foundation: Excels in multimodal reasoning, coding, agent tasks, and visual understanding benchmarks.
  • Efficient Hybrid Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference.
  • Scalable RL Generalization: Demonstrates robust real-world adaptability through reinforcement learning across diverse environments.
  • Multilingual Support: Expanded linguistic coverage to 201 languages and dialects.
  • Ultra-Long Context: Natively supports 262,144 tokens, with extensibility to over 1 million tokens.

Good For

  • Applications requiring a less restrictive, decensored language model.
  • Complex multimodal tasks involving both text and visual inputs, including video understanding.
  • Agentic applications and tool calling, with strong performance in general agent and search agent benchmarks.
  • Code generation and problem-solving, as indicated by strong performance on coding benchmarks like SWE-bench Verified and LiveCodeBench.
  • Scenarios demanding extensive context understanding, such as document analysis or long-form content generation.

Finetunes
1 models