Skip to content

SouthwardTech

Technology, industry and investment across China and the Global South

5 projects tracked1 Contracted4 AnnouncedReviewed

AI

Abu Dhabi’s AI lab builds new Arabic models on its own base

Jais 3 is built on K2 Horizon, a model family MBZUAI's research institute says it trained itself, just over a year after the institute built a reasoning model on Alibaba's Qwen

An Abu Dhabi research institute has released Jais 3, a family of six open Arabic language models built on a base model it says it trained itself.

The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Inception42, part of the G42 group, released the models on 15 October 2026 in collaboration with Cerebras Systems, according to a joint statement. IFM calls Jais 3 “a fully sovereign model for the UAE”, saying every model in the chain, from the pre-training of its K2 Horizon base to Jais 3’s final training stage, was trained by the institute.

The models range from 0.9 billion to 375 billion parameters, cover Modern Standard Arabic and dialects such as Emirati and Egyptian, and are built for multi-step tasks such as calling software tools. They are released under an Apache 2.0 licence.

Training a base model is a step many national AI efforts skip. Of the more than 170 sovereign language models tracked by Counterpoint Research across over 55 countries, 56 per cent are adapted from existing open models and 44 per cent were built from the ground up, according to analysis from the firm’s H1 2026 Sovereign AI LLM Index published on 21 September 2026. Meta’s Llama underpins 38 per cent of the adapted models and Alibaba’s Qwen 20 per cent. Egypt’s Karnak-40B, developed by the Applied Innovation Center, an affiliate of the country’s communications ministry, is built on Alibaba’s Qwen3-30B-A3B, according to its model card.

MBZUAI has used both routes. Its LLM360 project released K2-65B, trained from scratch, in May 2024. But K2 Think, a 32-billion-parameter reasoning model IFM launched with G42 on 9 September 2025, was built on Alibaba’s Qwen2.5-32B, according to its technical paper and Hugging Face listing. In December 2025 IFM released K2-V2, a 70-billion-parameter base model it says it trained from scratch, and in January 2026 it released K2 Think V2 on top of it. On 3 September 2026 it released K2 Horizon, a new family of six base models. The Jais line has also moved between the two routes: the first Jais models, released in 2023, were trained from scratch; Inception built Jais 70B on Meta’s Llama 2 in 2024; and Jais 2, released in December 2025, was again trained from scratch.

Timeline of base models behind MBZUAI's K2 and Jais lines from August 2023 to October 2026. K2 Think (September 2025) and Jais 70B (August 2024) were built on other labs' open models, Alibaba's Qwen2.5 and Meta's Llama 2; Jais 13B, K2-65B, K2-V2, Jais 2 and K2 Horizon were trained in-house, and Jais 3 (October 2026) is built on K2 Horizon.
Sources: Hugging Face model cards; arXiv; MBZUAI, IFM and G42 announcements

Open model code and designs are routinely reused; the costly part is training the weights, the billions of values a model learns. Jais 3’s published code carries a header crediting Alibaba’s Qwen team and Hugging Face, and its 7-billion-parameter version shares Qwen3-8B‘s number of layers, width and attention heads while using its own vocabulary.

About 59 per cent of the 151 billion Arabic tokens used in Jais 3’s continued pre-training, the stage that adapts K2 Horizon to Arabic, were synthetic or machine-translated, according to IFM. For its later fine-tuning data, the institute said “a strong open-weight model” translated conversations and generated agent training data, without naming the model. K2 Horizon’s technical report, which its model card said was expected at the end of September 2026, had not been published by 20 October 2026.

On IFM’s own evaluation, the 7-billion-parameter Jais 3 averaged 79.4 on Arabic tasks, ahead of the 70-billion-parameter Jais 2 at 66.7. The scores could not be independently verified. On agentic tasks, Jais 3 led every open model IFM tested at the 3.7-billion and 7-billion sizes; at 32 billion and 375 billion, the only open models it reported scoring higher were Alibaba’s Qwen3.8.

MBZUAI and Inception42 have not named any users of Jais 3. The models are available on Hugging Face, with support for the vLLM and SGLang serving frameworks, through an API on request and a chat interface; a Jais agentic platform is listed as “coming soon”.

Published 20 October 2026 · Last updated 20 October 2026 · Corrections to corrections@southwardtech.com

Weekly newsletter

The Southward Brief

A weekly briefing on the technologies, companies and industrial shifts reshaping China and the Global South.

Required. Enter a valid email address.

By subscribing, you agree to receive The Southward Brief. You can unsubscribe at any time. Read our Privacy Policy.