96kevinli29/Qwen3-4B-Multilingual-Reasoning

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 1, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The 96kevinli29/Qwen3-4B-Multilingual-Reasoning model is a 4 billion parameter Qwen3-based causal language model with a 32768 token context length, specifically fine-tuned for multilingual mathematical reasoning. Unlike standard models that reason in English and translate, this model generates its chain-of-thought in the target language across English, Chinese, Spanish, French, Arabic, and Russian. It excels at providing transparent, language-consistent reasoning traces for math problems in these six languages.

Loading preview...

Multilingual Reasoning with Qwen3-4B

This model, 96kevinli29/Qwen3-4B-Multilingual-Reasoning, is a specialized version of the Qwen3-4B architecture, fine-tuned to perform multilingual chain-of-thought reasoning in six languages: English, Chinese, Spanish, French, Arabic, and Russian. Its key differentiator is the ability to produce its internal <think>...</think> reasoning trace natively in the target language, rather than thinking in English and then translating only the final answer. This provides transparent and language-consistent mathematical problem-solving.

Key Capabilities

  • Native Multilingual CoT: Generates step-by-step reasoning directly in the specified language.
  • 6-Language Support: Covers English, Chinese, Spanish, French, Arabic, and Russian.
  • Mathematical Reasoning: Fine-tuned on a balanced 6-language long chain-of-thought math dataset (SFT-Math-45k-mling6).
  • Qwen3-4B Base: Built upon the official Qwen/Qwen3-4B post-trained model.

Good For

  • Applications requiring transparent, step-by-step mathematical reasoning in any of the supported languages.
  • Educational tools or systems where understanding the reasoning process in a specific language is crucial.
  • Research into multilingual reasoning and language consistency in LLMs.

Limitations

  • Some scripts, particularly Arabic and Russian, may exhibit generation-stability issues (looping or not terminating) during in-language reasoning. Users should consider both accuracy and generation-completion rates.

The model's training data was derived from translating a subset of OpenR1-Math using Qwen3-14B to preserve fidelity and formatting.