Open sourceModel launchArabic

Jais 3 launches as six open Arabic-native models, from 0.9B to 375B, plus Jais Speech

Institute of Foundation Models ·

Bar charts from the Jais-3-375B-A23B model card: on Arabic general knowledge it scores 92.4, tied with Qwen3.8-Flash-Next, and on Arabic STEM and math it scores 93.6, behind MiniMax M3, NVIDIA Nemotron 3 Ultra and two Qwen models

MBZUAI's Institute of Foundation Models and Inception42 released Jais 3 on Oct 15: six open language models from 900M to 375B parameters (the flagship is a mixture-of-experts model running 23B per token), all with a 512K context window, plus Jais Speech for Arabic speech recognition across dialects. Built on IFM's K2 Horizon models, they're on Hugging Face under Apache-2.0 with vLLM and SGLang support. In IFM's own Arabic tests, the 7B model's average rises from 56.4 for its K2 Horizon base to 79.6. IFM is also releasing the benchmarks and code it used.

How you’d use it: Run an open model that writes, transcribes and follows instructions in Arabic dialects, sized to your hardware.

Sources

← All AI news

Get the daily AI news

The day's important AI news by email: short, plain summaries, each linked to its source. Unsubscribe any time.

Prefer email? Write to [email protected] ([email protected]).