ConicCat/Qwen3.5-27B-Writer-V2

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

ConicCat/Qwen3.5-27B-Writer-V2 is a 27 billion parameter Qwen3.5-based language model developed by ConicCat, fine-tuned for high-quality writing and roleplay across various domains. It utilizes a curriculum learning approach, starting with roleplay data and progressing to writing data, to enhance generalization. The model emphasizes writing quality and maintains general instruction-following capabilities through a diverse training dataset, supporting a 32768 token context length.

Loading preview...

ConicCat/Qwen3.5-27B-Writer-V2 Overview

ConicCat/Qwen3.5-27B-Writer-V2 is a 27 billion parameter model built on the Qwen3.5 architecture, specifically fine-tuned for superior writing quality and roleplay performance. This model employs a unique curriculum learning strategy, initially training on roleplay data before advancing to higher-quality writing data, which helps it generalize effectively across both domains.

Key Capabilities & Training

  • Writing & Roleplay Focus: The primary emphasis is on generating high-quality text for both creative writing and roleplay scenarios, with strong generalization.
  • Curriculum Learning: Training began with lower-quality roleplay data, followed by extensive training on higher-quality writing data (book chunks) to refine its output.
  • Instruction Following: Despite its specialized focus, the model incorporates instruct data (from internlm/Condor-SFT-20K) to preserve general intellect and instruction-following abilities.
  • Repetition Mitigation: Utilizes the ConicCat/AntiRep dataset to reduce repetitive outputs.
  • Diverse Datasets: Trained on a mixture including ConicCat/AntiRep, internlm/Condor-SFT-20K, ConicCat/Gutenberg-SFT (reformatted), and ConicCat/MiniC2_V3.2 (cleaned and updated).

Recommended Usage

  • Chat Template: Best used with the ChatML template, optionally with <think>\n\n</think>\n or <think>\n prefill.
  • Parameters: Recommended settings include temperature = 0.7, top_p = 0.95, and a moderate dry penalty of 0.4-0.8.
  • Resource Efficiency: Quantized versions like Q4_K_M can run with approximately 100k context on 24GB VRAM, while IQ4_XS may fit on 16GB VRAM for 20-24k context.