At Misraj, we view bilingual Arabic-English translation as a foundational capability for modern AI systems across education, research, government platforms, and digital applications. While large translation models exist, they are resource-heavy and often fail to fully capture the cultural and linguistic nuances present in Arabic.

Mutarjim is our efficient, lightweight translation model designed to offer high-quality Arabic-English and English-Arabic translations without requiring expensive hardware. Despite its compact size, it delivers competitive performance compared to much larger models.

The problem Mutarjim solves

Most translators struggle with:

1. Deep semantic and cultural understanding of Arabic.

2. High computational cost.

3. Lack of a comprehensive bilingual benchmark.

Mutarjim directly addresses these issues.

How we built and trained Mutarjim

Our dataset spans:

• academic text

• governmental documents

• cultural and literary content

• technical and scientific material

• conversational data

We apply strict filtering and quality control to ensure accurate and natural bilingual pairs.

Tarjama-25: A benchmark introduced by Misraj

Tarjama-25 is an open, domain-rich benchmark covering law, medicine, education, governance, media, and technical writing, enabling transparent evaluation for all Arabic-English translation systems.

Results Table: Mutarjim vs Other Translation Models

 

Summary:

Mutarjim provides the best blend of quality, speed, and efficiency. Its translations preserve both meaning and style better than many generic translation systems.

Conclusion

Mutarjim marks an important step in advancing Arabic-English translation. With its lightweight architecture and the introduction of Tarjama-25, it sets a new standard for fair, transparent, and domain-rich evaluation of translation systems.

Research link:

https://arxiv.org/html/2505.17894v1