chartreuse-verte/prose-rewriter-4b-v1.6

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 4, 2026License:agpl-3.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The chartreuse-verte/prose-rewriter-4b-v1.6 is a 4 billion parameter Qwen3-4B-Base derivative model, fine-tuned as a paragraph-level prose rewriter. It specializes in transforming LLM-generated text to sound more human while preserving original semantics. This model is particularly optimized for rewriting fictional prose, offering three length transformation modes: match, inflate, and compress.

Loading preview...

Overview

prose-rewriter-4b-v1.6 is a 4 billion parameter model built upon Qwen/Qwen3-4B-Base, specifically designed to rewrite paragraph-level prose. Its primary function is to make text generated by large language models sound more human, ensuring the original meaning is maintained. This version, a successor to v1.4, features improved data cleaning and a gate to reduce invention, resulting in less hallucination and more truncation.

Key Capabilities

  • Prose Rewriting: Transforms LLM-generated text into more natural, human-like prose.
  • Semantic Preservation: Aims to retain the original meaning of the input text.
  • Length Transformation Modes: Supports match (rewrite in place), inflate (cut padded text), and compress (expand flattened text) via a mandatory edit block in the prompt.
  • Reduced Invention: Compared to its predecessor, v1.6 invents less and truncates more, improving fidelity.
  • Optimized for Fictional Prose: Best suited for narrative registers, including third and first-person fiction and dialogue.

Usage Notes

  • Prompt Format: Requires a specific chat template with source and edit roles. The edit block is crucial and dictates the length transformation.
  • Input Length: Optimal performance for inputs around 15 to 70 words. Inputs below 15 words may lead to fabrication.
  • Limitations: Not an instruct model, works on one paragraph per call, and is not designed for technical documentation. It will not pass AI detectors and is English-only.

Finetunes
1 models