hotdogs/Qwen3.8-27B-thinkingcap-abliterated

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

hotdogs/Qwen3.8-27B-thinkingcap-abliterated is a 27 billion parameter Qwen3.8-27B model, fine-tuned by hotdogs using a LoRA merge from the ThinkingCap efficient-reasoning dataset. This model is specifically designed to prevent 'runaway' thinking on complex, multi-step problems, significantly reducing output length without sacrificing accuracy. It excels at efficient reasoning by acting as 'runaway insurance' for hard problems, making it suitable for applications requiring concise and accurate responses.

Loading preview...

Overview

hotdogs/Qwen3.8-27B-thinkingcap-abliterated is a 27 billion parameter model based on Qwen3.8-27B, developed by hotdogs. It incorporates a LoRA merge from the ThinkingCap efficient-reasoning dataset, specifically checkpoint-42, epoch 1. The model is built upon an 'abliterated' base (hotdogs/Qwen3.8-27B-abliterated) which removes refusal behaviors, and then enhanced with the ThinkingCap layer.

Key Capabilities

  • Efficient Reasoning: The primary differentiator is its ability to prevent excessive 'runaway' thinking on hard, multi-step problems. Benchmarks show a 62% reduction in thinking length on runaway items compared to the base model.
  • Accuracy Preservation: Maintains identical accuracy on both easy (94.6%) and hard (70.0%) problem suites, demonstrating that efficiency gains do not come at the cost of correctness.
  • Refusal-Free: Inherits the refusal-removed characteristic from its abliterated base, ensuring direct answers.
  • Quantization Performance: Even in quantized GGUF versions (Q6_K, Q4_K_M), it performs strongly, scoring within 2.1 percentage points of the full BF16 Qwen3.8-27B on GPQA Diamond.

Good For

  • Applications requiring concise and accurate responses to complex problems.
  • Scenarios where controlling output verbosity and preventing model 'spiraling' is crucial.
  • Environments with hardware constraints, as its efficient reasoning can lead to shorter outputs and potentially faster inference, especially when combined with quantization.

Finetunes
2 models