HEBATRON: A Hebrew-Specialized Open-Weight Mixture-of-Experts Language Model
Paper • 2605.11255 • Published
🛡️ HEBATRON: Hebrew-Specialized Mamba2-MoE
HEBATRON is a state-of-the-art, high-performance language model specialized for the Hebrew language. Developed through a collaboration between PwC Israel, MAFAT, and AWS, it introduces a unique hybrid architecture combining Mamba2 and Mixture-of-Experts (MoE).
Technical Report: https://arxiv.org/abs/2605.11255
🚀 Model Summary
HEBATRON is designed to handle the structural and morphological complexities of Hebrew while providing linear scaling for long-context tasks. It is a localized and enhanced version of the Nemotron-3-Nano-30B framework, optimized for native-level reasoning in Hebrew and English.
📂 Technical Specifications
| Feature | Specification |
|---|---|
| Model Name | HEBATRON |
| Architecture | Hybrid Mamba2 (SSM) + Sparse MoE |
| Total Parameters | 31.6B |
| Active Parameters | ~3B per token |
| Context Window | 65,536 (64k) tokens |
| Hardware | NVIDIA Blackwell (B300) & H200 GPUs |
| Precision | FP8 Mixed-Precision |
🧬 Training Curriculum
The model was trained using a three-phase Curriculum Learning strategy:
- Phase 1: Formal Foundation (75.5B tokens) Focused on high-quality, structured Hebrew (legal, academic, and literary texts) to establish core grammatical rules.
- Phase 2: Colloquial Expansion (3.36B tokens) Integration of social media, forums, and informal web data to handle slang and modern registers.
- Phase 3: Long-Context Extension (20.4B tokens) Fine-tuning on dense, long-form documents to stabilize the 64k context window.
📊 Performance Evaluation
Hebrew Reasoning Benchmarks
- SNLI (Semantic Reasoning): 91.2% accuracy
- Israeli Trivia: 72.1% (+14pt vs base)
- Hebrew Average Reasoning: 73.8% (Surpassing DictaLM-3.0-Thinking)
- GSM8K (Math): 83.3% accuracy in native Hebrew
English Reasoning Benchmarks
- Psychometric Psi (EN): 91.6%
- English Reasoning Average: 86.0%
🎯 Intended Use & Limitations
- Intended Use: Advanced Hebrew document analysis, long-context summarization (legal/technical), and complex bilingual reasoning.
- Limitations: Users should verify outputs for factual accuracy as with any Large Language Model.
🤝 Credits
Project Leadership
- MAFAT Lead: Tal Geva (Project Lead), Matan Frank
- Technical Lead: Sarel Weinberger (PwC Next)
Core Teams
- PwC Israel Team: Noam Kayzer, Dan Revital, Ori Bar Joseph, Smadar Arbatz, Or Levi, Kate Zinkovskaia, Zevi Apini, Omer Baruch (PwC Next)
- MAFAT Team: Noam Ordan, Nadav Cordova
Partners & Collaborators
- Partners: Amir Nissan Hacohen (Origin.ai)
- Research Collaborators: Shaltiel Shmidman (Dicta), Mike Erlihson
- Infrastructure: Netanel Ilouz (AWS)
- Downloads last month
- 289