DictaLM-3.0-24B-Base: A New Frontier for Hebrew LLMs

DictaLM-3.0-24B-Base is a 24-billion-parameter foundational language model from Dicta, representing a significant advancement in Hebrew sovereign LLMs. This model was initialized from Mistral-Small-3.1-24B-Base-2503 and has been extensively trained on large Hebrew and English text corpora. It establishes a new state-of-the-art for its parameter class in Hebrew language performance, both as a base model and for subsequent chat model fine-tuning.

Key Capabilities & Features

• Bilingual Proficiency: Strong performance in both Hebrew and English, with a particular focus on Hebrew.
• State-of-the-Art Hebrew Performance: Achieves leading benchmarks for its size in Hebrew language tasks.
• Base Model Design: Intended for further fine-tuning to create specialized applications, including chat models.
• Full Precision: Available in BF16 precision.

Use Cases & Considerations

This model is ideal for developers and researchers looking to build applications that require robust Hebrew language understanding and generation. As a base model, it provides a powerful starting point for custom fine-tuning for specific tasks. Users should note that this is not an instruction-tuned chat model and lacks built-in moderation mechanisms, requiring developers to implement their own safety measures for downstream applications. For more details, refer to the technical report.