lightonai/Qwen3-8B-EN

Qwen3-8B-EN: English Reasoning Model

Qwen3-8B-EN is an 8 billion parameter model developed by lightonai, fine-tuned from Qwen/Qwen3-8B-Base. Its core distinction is its design as a native reasoning model that generates its complete chain-of-thought and final answer exclusively in English. This model was released in conjunction with the paper "Rethinking the Multilingual Reasoning Gap with Layer Swap."

Key Capabilities & Features

• Native English Reasoning: Produces detailed, step-by-step reasoning traces entirely in English before providing the final English answer.
• Optimized for Reasoning Tasks: Specifically trained to enhance performance on complex analytical and problem-solving benchmarks.
• Extensive Context Window: Supports a context length of 32,768 tokens, allowing for processing of lengthy inputs and complex problems.
• Robust Training: Underwent full SFT (Supervised Fine-Tuning) on approximately 10 billion tokens over 2 epochs, utilizing the English split of the lightonai/Dolci-Think-SFT-32B-Multilingual dataset.

Performance Highlights

Evaluated on English versions of various benchmarks, Qwen3-8B-EN demonstrates strong performance:

• MGSM-Rev2: 98.96%
• Global-MMLU-Lite: 81.72%
• GPQA-Diamond: 55.66%
• AIME 24/25: 62.89%
• HumanEvalPlus: 85.75%

These scores reflect mean accuracy, with an average of 77.00% across the listed benchmarks, showcasing its proficiency in mathematical reasoning, general knowledge, and code generation within an English context.

Ideal Use Cases

• Applications requiring explicit, step-by-step English reasoning.
• Tasks involving complex problem-solving, mathematical calculations, or code generation where a clear chain of thought is beneficial.
• Scenarios demanding high accuracy on English-centric reasoning benchmarks.