logo

Translation & Localization

Enterprise Arabic translation, not general-purpose machine translation

Good Arabic translation is either too coarse to publish or too heavy to run.

Standard multilingual models flatten Arabic's grammar, morphology and dialectal variation. Massive general-purpose LLMs translate better but are too heavy and too slow to run in production workflows at volume.

Mutarjim delivers state-of-the-art English-to-Arabic translation at 1.5B parameters - small enough to ship in production, beating GPT-4o mini on Tarjama-25. Lahjawi covers dialect-to-dialect and dialect-to-MSA translation across 15 Arabic dialects.

Why a small model is the point

  • State of the art at 1.5B parameters

    Mutarjim is built on Kuwain 1.5B, Kawn Lab's own Arabic-centric language model, using a two-phase training approach on a curated bilingual corpus. It is small enough to ship in production, and it beats GPT-4o mini on English-to-Arabic.

  • Dialect is in scope

    Lahjawi handles dialect-to-dialect and dialect-to-MSA translation across 15 Arabic dialects, which is where general multilingual models stop being useful entirely.

  • Benchmarked on something honest

    Tarjama-25 is 5,000 expert-reviewed sentence pairs across diverse domains and lengths, built specifically to fix the domain narrowness and English-source bias that make existing Arabic-English benchmarks unreliable.

What it translates

  • Bidirectional, long-form, domain-heavy

    Arabic-to-English and English-to-Arabic across medical, legal and technical content, preserving meaning through complex sentences rather than flattening them.

  • Fifteen Arabic dialects

    Dialect-to-dialect and dialect-to-Modern Standard Arabic, so regional content is translatable material rather than an exception.

  • Diacritization as a preprocessing step

    Sadeed resolves Arabic's default ambiguity by reading the grammar and meaning around each word. That is directly applicable to machine translation, search indexes and text-to-speech.

  • Sized for real workflows

    Small enough for real-time deployment and low-resource environments, where a large general LLM is too slow or too expensive to run at volume.

The numbers behind it

  • 1.5BParameters: Mutarjim rivals models 20× its size
  • SOTAEnglish-to-Arabic on Tarjama-25, outperforming GPT-4o mini
  • 15Arabic dialects covered by Lahjawi

What changes

  • Translation moves inside the workflow

    It runs as a step in the pipeline rather than as a separate vendor engagement with its own turnaround.

  • Dialect content stops being untranslatable

    Regional material becomes something the system can process rather than something routed to a human every time.

  • Ambiguity is resolved before it propagates

    Diacritized text means search, TTS and downstream NLP stop inheriting the guesswork that undiacritized Arabic hands them.

How it deploys

  • Cloud API

    Through Kawn Console: one account and one integration point for the whole model suite.

  • On-premises

    For organisations with data sovereignty requirements. The compact model size is what makes this practical rather than theoretical.

  • Low-resource environments

    Real-time deployment where a large general LLM would not fit the latency or cost budget.

Let's talk about what you're trying to build.

Tell us the problem. We'll tell you honestly whether AI is the right answer.

Translation & Localization | Misraj AI | Next-Gen Arabic AI Lab