Hebrew-LLMs

April 17, 2025 · View on GitHub

A curated collection of Hebrew language AI models available on Hugging Face as of April 17, 2025. This repository aims to provide a good starting point for anyone looking to work with Hebrew language AI models.

About Hebrew Language and AI

Hebrew is a Semitic language with approximately 9 million native speakers worldwide, primarily in Israel. Despite its relatively small speaker base, Hebrew presents several interesting characteristics for AI research:

Modern vs Biblical Hebrew: There are significant differences between Modern Hebrew and Biblical Hebrew, with specialized models developed for biblical text analysis.
Punctuation Challenges: Modern written Hebrew typically lacks extensive punctuation, creating a need for specialized models that can infer and add appropriate punctuation.
Technological Hub: Israel is a renowned center for technology and AI research, making Hebrew language AI models particularly interesting from an experimental and innovation perspective.
Rich Linguistic Structure: Hebrew's non-Latin script, right-to-left writing system, and complex morphology present unique challenges for language models.

These factors make Hebrew language AI development both challenging and valuable, with applications ranging from biblical text analysis to modern NLP tasks.

Hebrew-LLMs

Large Language Models (LLMs)

Mistral Fine-tunes

Model	Link
Hebrew-Mistral-7B
Hebrew-Mistral-7B-200K
Hebrew-Mistral-7B_Chat-GGUF
Hebrew-Mistral-7B-Instruct-v0.1-GGUF

Mixtral Fine-tunes

Model	Link
Hebrew-Mixtral-8x22B

Gemma/Google Fine-tunes

Model	Link
Hebrew-Gemma-11B
Hebrew-Gemma-11B-Instruct
Hebrew-Gemma-11B-V2-mlx-4bit

Niche Text Models

Summarization

Biblical

Metaphor Detection

Translation

Offensive Language

Punctuation

Sentiment Analysis

Specialized Language Models

Hebrew to SQL

Hebrew Medical Terms (NER)

ASR Models (Speech Recognition)

TTS Models (Text-to-Speech)

Note: This section will be populated with Hebrew TTS models in the future.

Benchmarks and Leaderboards

The Hebrew LLM Leaderboard provides valuable insights into the performance of various models on Hebrew language tasks:

Resource	Link
Hebrew LLM Leaderboard
Hebrew Question Answering Dataset

Leaderboard Insights

An interesting observation from the leaderboard is that large multilingual LLMs (like Mistral and Meta-Llama models) generally outperform specialized Hebrew models due to their significantly larger parameter counts. However, specialized Hebrew models still appear on the leaderboard and perform reasonably well considering their size constraints.

The benchmark evaluates models across several categories: