IFM and Inception42 release the six-model Jais 3 family

Background

Jais 3 is a family of Arabic-centric language models from the Institute of Foundation Models (IFM) and Inception42. It was not trained from an empty starting point: all six of its language models are built on K2 Horizon, the model fleet IFM had released the month before ^[1]. Building on a base model and adapting it is a common pattern, because most general capability — grammar, reasoning, world knowledge — is already in the base, and the adaptation step pushes the model toward a particular language and set of tasks. The size of that step is visible in the published numbers: at 7 billion parameters, the Arabic average rises from 56.4 for the K2 Horizon base model to 79.6 after adaptation ^[1].

What was announced

IFM and Inception42 announced Jais 3 on 15 October 2026 in Abu Dhabi ^[1]. The release covered six language models, from 900 million to 375 billion parameters, along with Jais Speech, a speech recogniser ^[1]. Cerebras Systems collaborated on the release ^[1].

Parameter count is a rough measure of a model's capacity, and also of how much computing power it takes to run; a range from 900 million to 375 billion spans models that fit into small footprints and models that need substantial infrastructure, giving users a choice along that trade-off.

How the models were evaluated

All six were evaluated on an audited suite of 24 Arabic benchmarks built by IFM and Inception42 ^[1]. Each model ships with five training checkpoints, and the benchmarks and evaluation code are published alongside ^[1]. A checkpoint is a saved intermediate state from training; releasing five of them lets a user pick a version at a different point in the adaptation process rather than accept only the final one. Publishing benchmarks and evaluation code with the weights means outside readers can rerun the comparison instead of taking the reported scores on trust.

Writing the test items in Arabic rather than translating them from English also changes what is measured. A translated benchmark tends to carry the grammar and cultural assumptions of the language it was written in; a suite written for Arabic can test what Arabic actually does.

Arabic data and speech

IFM built 7.1 billion tokens of Arabic long-context training data, about 17 percent of supervised fine-tuning ^[2]. Supervised fine-tuning is the stage where a model is trained on example inputs paired with desired outputs; the long-context part means the material includes long documents, so the model learns to keep track of information across far more text than a single paragraph.

Jais Speech covers Modern Standard Arabic and the dialects spoken across 19 Arab countries ^[1]. Modern Standard Arabic is the shared written and formal register, while everyday speech tends to sit closer to a regional dialect, so a recogniser that covers both is easier to put to work.

What comes next

Ashish Koshy, chief executive of Inception42, said that building an Arabic model is not a translation exercise ^[1]. That matches the shape of the release: the benchmarks were written for Arabic, and the speech model was built for both the standard form and 19 countries' dialects. A Jais agentic platform for everyday workflows was described as coming soon ^[1]. An agentic system is one meant to carry out multi-step tasks rather than answer a single prompt, which is a different kind of target from the scored benchmarks above.