rawmodels/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:Transformer Open Weights Featherless Exclusive Cold

rawmodels/Qwen3.8-27B-antislop is a 27 billion parameter Qwen3.8-based language model fine-tuned to produce more human-like, less "LLM-like" prose. It utilizes Reinforcement Learning from eXpert Aligned Rubrics (RL-XAR) to reduce generic or "sloppy" AI writing. The model retains the native vision-language capabilities of its base, including a vision tower and multi-token prediction head, and excels at generating natural, concrete text across various domains like tweets, scientific abstracts, and blog essays. It achieves a mean quality score of 8.28 (0-10) on held-out prose and ranks 4th on EQ-Bench 4 with an Elo of 1320.1.

Loading preview...

rawmodels/Qwen3.8-27B-antislop: Human-like Prose Generation

This model is a 27 billion parameter Qwen3.8 variant, specifically post-trained to generate text that sounds more human and less like typical LLM output. It achieves this by reimplementing Meta FAIR's Reinforcement Learning from eXpert Aligned Rubrics (RL-XAR) method, which uses expert-aligned rubrics as a reward signal during training. Unlike many models that can produce generic or "sloppy" prose, Qwen3.8-27B-antislop focuses on craft and naturalness.

Key Capabilities

  • Human-like Text Generation: Significantly reduces "AI slop" by producing concrete, in-voice, and less abstract phrasing, as demonstrated by examples in essay continuations, product memos, and tweets.
  • Retains Vision-Language Features: The base Qwen/Qwen3.8-27B's native vision tower and multi-token prediction head are preserved, allowing for potential multimodal applications.
  • Superior Prose Quality: Achieves a mean quality score of 8.28 (on a 0-10 scale) on held-out prose, outperforming the base Qwen3.8-27B (6.31) and Altworld/Hemmingway-1 (7.10).
  • High EQ-Bench Ranking: Ranks 4th overall on EQ-Bench 4 with an Elo score of 1320.1, placing it above models like GPT-5.5 and Claude Opus 4.7, indicating strong performance in open-ended conversation and naturalness.

Good for

  • Content Creation: Ideal for generating articles, blog posts, social media content, or any text where a natural, non-AI-generated tone is crucial.
  • Creative Writing: Useful for tasks requiring nuanced, engaging, and less formulaic prose.
  • Avoiding Generic Output: Developers seeking to mitigate the common issue of LLMs producing overly fluent but ultimately bland or predictable text.
  • Applications requiring authentic communication: Scenarios where the generated text needs to resonate as genuinely human, such as customer service responses, marketing copy, or narrative development.

Finetunes
7 models