Hebrew Math Tutor v1: Specialized Mathematical Reasoning in Hebrew

Intel's Hebrew Math Tutor v1 is a 4 billion parameter decoder-only causal language model, built upon Qwen3-4B-Thinking-2507. This model is specifically fine-tuned to provide detailed, step-by-step mathematical reasoning and solutions in Hebrew, making advanced AI math capabilities accessible for Hebrew-language education.

Key Capabilities & Features

• Hebrew Mathematical Reasoning: Delivers clear, pedagogically sound explanations for math problems entirely in Hebrew.
• High Accuracy: Maintains mathematical precision, showing improvements in pass@16 on Hebrew evaluations across benchmarks like MATH500 and AIME.
• Language Adaptation: Achieves 95-100% Hebrew output for answers, a significant increase (25-61.5%) compared to its base model.
• Multilingual Preservation: Retains underlying multilingual capabilities for broader applications.
• Supervised Fine-Tuning: Trained on ~10,000 translated problems from the OpenMathReasoning dataset, focusing on Hebrew SFT.

Ideal Use Cases

• Educational Technology: Developing Hebrew-language math tutoring systems and learning platforms.
• Research: Supporting mathematical reasoning research within Hebrew educational contexts.
• Accessibility: Providing advanced math AI assistance to Hebrew-speaking communities.
• Prototype Development: Building Hebrew-first educational AI applications.