Institute of Foundation Models (IFM)

What it is

The Institute of Foundation Models (IFM) is part of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi^[1]. It describes its mission as the open and independent development of frontier foundation models^[1].

A foundation model is a large model trained on broad data that is then adapted to many downstream tasks instead of being built for one narrow purpose. "Frontier" generally refers to the most capable models available at a given moment, and "open" to work whose methods or weights are published rather than reachable only through a service. The institute's stated mission places it in that territory: research on large models, framed as open and as independent^[1].

Where it came from

MBZUAI launched the institute in May 2025^[2][3]. The university's president, Eric Xing, is described as founder of the institute^[2]. MBZUAI itself was founded in 2019 and enrolled its first graduate students in 2021^[3]; according to its Jais 3 release it counts nearly a thousand students^[3]. The institute is therefore a new unit inside a university that is itself only a few years old — an organisation assembled from scratch rather than a lab grown out of an older one.

Three sites

The institute works from three sites: Abu Dhabi, Silicon Valley and Paris^[2]. The Abu Dhabi office is on the MBZUAI campus in Masdar City^[4]. The Silicon Valley lab is at 150 Mathilda Place in Sunnyvale, California^[4] and is directed by Hector Liu^[2]. The Paris office is at 42 Rue Notre Dame des Victoires^[4]. Richard Morton is vice president and executive director of the institute^[3].

Splitting a research organisation across several cities is a common way to recruit people who would rather not relocate and to keep a foothold in more than one local research community; the cost is coordination. For this institute the arrangement is explicit — three named sites, each with its own address and, for the Silicon Valley lab, its own director^[2][4].

The model portfolio

Its portfolio covers the K2 series, the Arabic Jais line and PAN, described as a world model for embodied reasoning^[2][3]. The three lines point in different directions. Jais is an Arabic-focused line, K2 a general series of models.

A world model, in the usual sense of the term, learns to predict how an environment changes over time, so that a system can plan ahead rather than only react to what it observes. "Embodied" reasoning means reasoning carried out by an agent that acts in a physical or simulated environment — the kind of capability that matters for robots and for software that has to operate somewhere rather than just answer questions. PAN is presented in those terms^[2][3].

Where it sits

What distinguishes the institute's stated position is the pairing of an open, independent framing with an unusual geographic footprint: a base at a Gulf university, plus labs in the United States and Europe^[1][2]. It enters a field where frontier models are usually built either inside large technology companies or inside a small number of well-established academic groups, and it argues that the work can be done openly and independently instead^[1].

The concrete output to judge it by is the model line itself: K2, Jais and PAN, and what the institute publishes about how they are built and who can use them^[2][3].