IstroSec/ThinkingCap-Qwen3.8-27B-abliterated

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

IstroSec/ThinkingCap-Qwen3.8-27B-abliterated is a 27 billion parameter language model, derived from bottlecapai's Qwen3.8-27B, with a 32768 token context length. This model has been modified to significantly reduce refusal behavior by merging an abliteration adapter, while retaining the original's calibrated thinking length, vision capabilities, and Multi-Token Prediction (MTP) head for speculative decoding. It is primarily designed for research, red-teaming, and creative applications where an uncensored model is required, offering greater compliance with diverse user requests.

Loading preview...

Overview

IstroSec/ThinkingCap-Qwen3.8-27B-abliterated is a 27 billion parameter language model based on bottlecapai's ThinkingCap-Qwen3.8-27B. Its primary differentiator is the removal of refusal behavior, achieved by merging a published abliteration adapter from MuXodious. This modification significantly reduces the model's tendency to decline requests, making it suitable for use cases requiring an uncensored output. The model retains the original ThinkingCap's calibrated thinking length, vision tower, and Multi-Token Prediction (MTP) head for speculative decoding, ensuring its core functionalities remain intact.

Key Capabilities

  • Reduced Refusal Behavior: Evaluated to refuse only 6 out of 100 harmful prompts, a substantial reduction from the original model's 97 refusals.
  • Calibrated Thinking Length: Preserves the reasoning_effort control (xhigh, medium, low) for managing the model's thought process.
  • Vision Capabilities: The integrated Vision Transformer (ViT) tower and merger are untouched, allowing for multimodal inputs.
  • Speculative Decoding (MTP): Includes an unmodified MTP head, supporting efficient generation with methods like vLLM's speculative decoding.
  • Tool Calling: Retains the original chat template and parser flags for tool-use functionality.

Good For

  • Research and Red-Teaming: Ideal for exploring model boundaries and testing safety mechanisms without inherent refusal.
  • Creative Work: Suitable for applications where unconstrained text generation is desired.
  • Agent Pipelines: Useful in scenarios where over-refusal from a language model is problematic.
  • Custom Safety Layers: Designed to be run behind user-defined policy controls, placing the safety responsibility on the implementer.

Finetunes
6 models