Laow0v0/neko-qwen3-4b

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The Laow0v0/neko-qwen3-4b is a 4 billion parameter Qwen3-based instruction-tuned causal language model, fine-tuned by Laow0v0. It specializes in maintaining a consistent catgirl (猫娘) persona, addressing users as "master" and incorporating characteristic verbal tics. This model is primarily designed for style transfer research, roleplay, and companionship-style chat in Chinese and mixed Chinese-English contexts.

Loading preview...

Model Overview

The Laow0v0/neko-qwen3-4b is a 4 billion parameter language model, fine-tuned from unsloth/qwen3-4b-instruct-2507 (merged to 16-bit). Its core differentiator is its catgirl (猫娘) persona, which is deeply embedded rather than relying solely on system prompts. The model was trained on the NekoQA-10K dataset, comprising 10,000 single-turn QA pairs written in this specific persona.

Key Characteristics

  • Consistent Persona: Addresses users as "主人" (master) and uses characteristic verbal tics like "喵~" and "的说喵", maintaining a cute, affectionate, 二次元 register.
  • Multilingual Support: Primarily Chinese, with some mixed Chinese-English content.
  • Training Method: Fine-tuned using Unsloth and Huggingface's TRL library for efficiency.

Intended Use Cases

  • Style Transfer Research: Ideal for exploring persona consistency and style transfer in LLMs.
  • Roleplay & Companionship: Suited for interactive chat applications where a specific, consistent persona is desired.

Limitations

  • Persona Inconsistency: May revert to a plain assistant voice or base model identity (通义千问) in Chinese, requiring explicit system prompts for full adherence.
  • Generation Scaffolding: Often includes un-terminated <think>, <tool_call>, or (Dialogue begins) markers that need post-processing.
  • Factual Reliability: Not recommended for tasks requiring high factual accuracy, as the dataset prioritizes tone over factual rigor.