Jais 3: open Arabic models for speech, writing and agents

What Jais 3 is

Jais 3 is built on K2 Horizon, IFM's fully open model fleet released in September, and is its first adaptation to a new language. All six models share one training recipe and agent capabilities, with a 512,000-token context window (PR Newswire).

The models handle Modern Standard Arabic and regional dialects from the Gulf to North Africa. Jais Speech transcribes dialects from 19 Arab countries in noisy conditions and keeps mid-sentence English or French in its original script.

What the numbers say

The developers tested on their own audited suite of 24 Arabic benchmarks. They say the largest model matches the best open models on Arabic understanding and leads them on agentic tasks. At 7 billion parameters the average score rose from 56.4 for the K2 Horizon base to 79.6.

Jais Speech posted the lowest average word and character error rates across six public test sets among the models compared, including commercial services. These are the developers' own results; independent checks are still pending.

Why it matters

More than 400 million people speak Arabic, yet general-purpose models often drift between dialects and registers. Jais 3 shows how an open foundation model can be adapted to a new language and culture without starting from scratch, a useful template for other mid-sized languages.

IFM also released the benchmarks and evaluation code, plus five training checkpoints per model. Jais 3 and Jais Speech are on Hugging Face with vLLM and SGLang support.