AlexWortega/Qwen3.8-27B-antislop

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

AlexWortega/Qwen3.8-27B-antislop is a 27 billion parameter language model, post-trained from Qwen3.8-27B, designed to produce text that reads more human-like and less like typical LLM output. It utilizes a novel Reinforcement Learning from eXpert Aligned Rubrics (RL-XAR) approach, inspired by Meta FAIR's 'unslop' work, to reduce generic 'AI slop'. This model excels at generating prose with a consistent, concrete voice across various domains like tweets, articles, and blogs, making it suitable for applications requiring natural, high-quality text generation.

Loading preview...

Qwen3.8-27B-antislop: Reducing AI Slop for Human-like Text

AlexWortega/Qwen3.8-27B-antislop is a 27 billion parameter model, fine-tuned from Qwen3.8-27B, specifically engineered to generate text that mimics human writing style, moving away from the often generic and abstract output characteristic of many LLMs. This model achieves its unique quality by implementing a sophisticated Reinforcement Learning from eXpert Aligned Rubrics (RL-XAR) method, a reinterpretation of Meta FAIR's "unslop" research.

Key Capabilities

  • Human-like Prose Generation: Significantly reduces "AI slop" by producing concrete, in-voice text across diverse content types.
  • High-Quality Output: Achieves a mean quality score of 8.28 (0-10 scale) on held-out prose, outperforming the base Qwen3.8-27B and other creative-writing finetunes.
  • Strong Benchmarking: Ranks #4 on EQ-Bench 4 with an Elo score of 1320.1, placing it above models like GPT-5.5 and Claude Opus 4.7/4.8 in human-preference evaluations.
  • Contextual Consistency: Maintains a consistent voice and avoids abstract, clichéd phrasing, as demonstrated in examples for continuing essays, product memos, and tweets.
  • RL-XAR Methodology: Utilizes a calibrated Claude Opus judge with length-blind, frozen per-item rubrics for reward signaling during GRPO training, ensuring quality over mere fluency.

Good For

  • Content Creation: Ideal for generating blog posts, articles, social media updates, and other written content where a natural, non-AI-generated feel is crucial.
  • Creative Writing: Suitable for tasks requiring a distinct and consistent authorial voice, such as continuing narratives or crafting specific tones.
  • Avoiding Generic Output: Excellent for applications where the goal is to produce text that stands out from typical LLM responses, offering more concrete and engaging prose.
  • Research and Development: Provides a strong baseline for further experimentation in reducing AI-generated artifacts and improving text naturalness.

Finetunes
3 models