NewstaR/Newstar-Qwen3-0.6B

Overview

Newstar-Qwen3-0.6B is a finetuned version of the Qwen3-0.6B base model. It uses Newstar’s instruction tuning on top of Qwen3’s pretrained weights. The tuning was done using the ITP-v2 dataset.

This model is designed without thinking capabilities. It intentionally avoids the reasoning mode that Qwen3 supports. Its primary purpose is to offer a Qwen3-based model that focuses on straightforward instruction following without engaging in complex reasoning.

Model Details

• —Base model: Qwen3-0.6B (causal language model)
• —Finetuning: Instruction tuning by Newstar on ITP-v2 dataset
• —Parameters: 0.6 billion
• —Layers: 28
• —Attention heads: 16 Q, 8 KV (GQA)
• —Context length: 32,768 tokens

Intended Use

• —General instruction following without reasoning or "thinking"
• —Simple and efficient dialogue or text generation tasks
• —Scenarios where disabling complex logical or mathematical reasoning is preferred

Performance and Limitations

• —Benchmarks: Not yet tested
• —Informally observed to be less capable than Qwen3 Instruct in reasoning and complex tasks
• —Lacks the advanced thinking mode of the original Qwen3 model
• —Use when reasoning is not required or to avoid thinking mode overhead

How It Differs from Qwen3-0.6B

• —Qwen3 supports both thinking and non-thinking modes
• —Newstar-Qwen3-0.6B disables thinking mode completely
• —Less capable in math, coding, and logic but simpler for basic instructions

Usage Notes

• —Thinking mode (enable_thinking=True) is disabled
• —Model does not generate <think>...</think> reasoning blocks
• —Recommended generation settings (non-thinking mode):
• —Temperature: 0.7
• —Top-p: 0.95
• —Top-k: 64
• —Rep-penalty: 1.03

Citations

@misc{qwen3technicalreport,
  title={Qwen3 Technical Report}, 
  author={Qwen Team},
  year={2025},
  eprint={2505.09388},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2505.09388}, 
}