croll83/Qwopus3.6-27B-v1-Abliterated-preview

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The croll83/Qwopus3.6-27B-v1-Abliterated-preview is an uncensored 27 billion parameter multimodal language model based on the Qwen3.6-27B architecture, featuring a 262K token context length and a vision encoder. This model is a Claude-distilled reasoning fine-tune of the original Qwen model, with its refusal direction removed using orthogonal-projection abliteration. It is designed for research and experimental use, excelling in structured reasoning and consistent answer styles across various tasks, including agentic coding and multimodal understanding.

Loading preview...

Qwopus3.6-27B-v1-Abliterated-preview: Uncensored Multimodal Reasoning

This model, developed by croll83, is an uncensored preview of the Jackrong/Qwopus3.6-27B-v1-preview, built upon the Qwen/Qwen3.6-27B architecture. It features approximately 28 billion parameters (plus a 440M vision encoder) and a substantial 262,144 token context length, with extensibility up to 1,010,000 tokens via YaRN scaling. The key differentiator is its "abliterated" nature, meaning the refusal direction in its activation space has been removed using orthogonal-projection, making it suitable for research into less constrained AI behavior.

Key Capabilities & Features

  • Uncensored Output: Abliteration removes safety filtering for research and experimental use.
  • Multimodal: Supports both text and vision inputs, including video, with a dedicated vision encoder.
  • Enhanced Reasoning: Fine-tuned for stronger reasoning quality, structured answers, and consistent style, derived from Claude distillation.
  • Agentic Coding: Excels in agentic coding tasks, frontend workflows, and repository-level reasoning.
  • Ultra-Long Context: Natively supports 262K tokens, extendable to over 1M tokens with YaRN.
  • Optimized Runtimes: Provides GGUF quantizations (Q4_K_M, TQ3_4S, NVFP4) and specific optimizations for llama.cpp-dgx on Blackwell-class GPUs, enabling high decode performance (e.g., 68.7 tok/s for JSON on GB10).

Ideal Use Cases

  • Research & Experimentation: Specifically designed for exploring model behavior without safety constraints.
  • Complex Reasoning Tasks: Suited for problems requiring structured thought processes and consistent output.
  • Multimodal Applications: Effective for tasks involving image, video, and text understanding.
  • Agent Development: Strong performance in agentic coding and tool-use scenarios.
  • High-Throughput Inference: Optimized for various inference frameworks like vLLM, SGLang, and llama.cpp for efficient deployment.

Finetunes
1 models