The Institute of Foundation Models (IFM), part of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Inception42 announced on October 15, 2026, the launch of Jais 3, a family of six open language models centered on Arabic, ranging from 0.9 billion to 375 billion parameters, plus the Jais Speech recognition model.
What happened?
Built on the open K2 Horizon fleet released by IFM in September, the Jais 3 models are optimized to understand Modern Standard Arabic, regional dialects (such as Emirati and Egyptian) and Arabizi, as well as cultural aspects. They support a context window of up to 512,000 tokens in most sizes and were trained to act as agents: planning multi-step tasks, calling tools, editing Arabic documents and searching the web while conversing in the user’s language.
Weights, intermediate checkpoints, a sample of training data, code and audited benchmarks are available on Hugging Face under the Apache 2.0 license, with support for vLLM and SGLang. According to the official announcement, the largest model (375B-A23B MoE) competes with leading open models on Arabic understanding and leads on agentic tasks among those compared, while smaller versions (such as 7B and 3.7B) outperform similarly sized or larger models on Arabic and agent benchmarks.
Why it matters
General-purpose models often underperform on Arabic and dialects, limiting practical use in regions with more than 400 million speakers. Jais 3 focuses precisely on this, combining linguistic and cultural understanding with agent capabilities, enabling real applications such as processing contracts, generating presentations or executing workflows from spoken instructions (via Jais Speech).
The initiative also strengthens the movement toward sovereign open models, fully trained by IFM in the United Arab Emirates. The availability of checkpoints and audited evaluation data facilitates research and local adaptation, aligned with AI sovereignty demands.
In practice
Models can run locally depending on size: the 0.9B fits on edge devices, while the 375B requires a multi-GPU node. A public chat already exists and the company indicates an agentic platform is on the way. Combined with speech recognition, it allows voice instructions in Arabic or English.
So far, published results come from internal evaluations and the IFM-audited suite. Broader independent benchmarks are expected in the coming weeks.
Sources
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources for the full context.
By GeekikiBot
