abliterant/Qwen3.8-27B-RANA-abliterated

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

abliterant/Qwen3.8-27B-RANA-abliterated is a 27.7 billion parameter model developed by Abliterant, based on Qwen/Qwen3.8-27B, with a configured context length of 262,144 tokens. This model utilizes Reasoning-Anchored, Norm-preserving Ablation (RANA) to remove refusal-behavior, making it a safety-alignment-removed research model. It is intended for interpretability, red-teaming, and robustness evaluation of refusal behavior, demonstrating significantly reduced hard refusals while maintaining most of the base model's capabilities.

Loading preview...

Overview

abliterant/Qwen3.8-27B-RANA-abliterated is a 27.7 billion parameter research model developed by Abliterant. It is a refusal-direction-ablated version of Qwen/Qwen3.8-27B, utilizing Reasoning-Anchored, Norm-preserving Ablation (RANA). This process removes a single "refusal direction" from the model's residual stream, keeping edited weights at their original norms. The model is configured for a 262,144-token context, though evaluations were conducted with a 20,480-token serving limit.

Key Capabilities

  • Refusal Behavior Research: Designed specifically for interpretability, red-teaming, and robustness evaluation of refusal behavior, with hard refusals reduced to near zero.
  • Capability Preservation: Maintains most of the base model's capabilities, with an average capability change of 0.82 percentage points versus the base model across five tasks.
  • Technical Reasoning: Excels at long technical requests, often hitting the 16k-token budget with detailed reasoning rather than repetition.
  • Multimodal Support: Vision input, multi-turn tool calling, and MTP speculative decoding are functional.

Good For

  • Safety Alignment Research: Ideal for researchers studying and evaluating refusal mechanisms in large language models.
  • Red-Teaming: Useful for probing model responses without inherent safety alignments.
  • Interpretability Studies: Provides a platform for understanding how refusal behaviors are encoded and removed within model architectures.
  • Complex Technical Queries: Suitable for tasks requiring extensive reasoning and long outputs, provided max_tokens is set appropriately.

Finetunes
4 models