Abu Dhabi IFM Releases Six Fully Open K2 Horizon AI Models

TL;DR: Abu Dhabi's Institute of Foundation Models released six K2 Horizon AI models with weights, training data, code and methods on September 3, 2026.

The Institute of Foundation Models (IFM) introduced a family of six foundation models called K2 Horizon on September 3, 2026, publishing weights, training data, program code and development methods alongside the systems, according to Reuters in its report titled "Abu Dhabi AI institute releases fully open-source models with training data, code | Reuters". The Abu Dhabi-based institute said researchers could use the package to examine, rebuild and modify the work, a level of disclosure that goes further than download-only releases. Founder Eric Xing told Reuters the bundle also contains intermediate checkpoints, so outside teams can trace construction steps and repeat tests.

The announcement centers on reproducibility rather than a single download. IFM described the collection as spanning 0.9 billion to 375 billion parameters, with shared architecture, training approach, interfaces and deployment tools, according to the Institute of Foundation Models press release titled "K2 Horizon Press Release | Institute of Foundation Models". A smaller vocabulary is used for the 0.9B system, the institute said. The organization also described a dynamic model routing technique that sends work to the most cost-effective member of the family, giving developers a route from early prototype to production without leaving the fleet.

What IFM published on September 3

IFM presented K2 Horizon as fully open, covering model weights, code, training data and methodologies, according to the Institute of Foundation Models press release. The institute said the materials let researchers and developers inspect, reproduce and adapt the models for their own projects.

Reuters reported that the release included weights, training data, code, methodologies and intermediate checkpoints. Xing told Reuters that access to those artifacts allows researchers to retrace the development process and reproduce results.

The institute framed the move as support for verification. Every model ships with its training data, recipe and evaluations, Xing said in the IFM announcement. That package, he said, reflects open science.

Why IFM says this goes beyond open weights

The distinction matters because many public model releases share downloadable weights while withholding information about construction. Reuters reported that the K2 Horizon approach goes beyond the "open-weight" method used by some Chinese developers, who make models available for download but offer limited insight into how they were built.

Reuters also contrasted the release with the more closed development practices of companies such as OpenAI and Anthropic, which neither release their models for download nor disclose the data and techniques behind them.

Xing told Reuters the objective was to establish a reference point for what a truly open model release can look like. He said the release was also intended to show policymakers, regulators and public-interest advocates that openness and competitive performance are not mutually exclusive, according to Reuters.

The press announcement used similar language, saying fully open code, training data and recipes represent a significant step forward in transparency, well beyond the "open weights" dialogue that has dominated industry headlines over recent weeks.

Six sizes from watches to enterprise servers

The fleet covers a wide span of deployment targets. IFM said the six models share core architecture, vocabulary, training methodology, interfaces and deployment tooling, with the exception of a smaller vocabulary for the 0.9B model.

The smallest system, at 0.9B parameters, is designed for highly constrained environments such as watches and glasses, according to the IFM announcement. Reuters described that end of the range as one designed for smartwatches and other constrained devices.

The 3.7B and 7B models are aimed at phones and other on-device applications, IFM said. The organization said the 3.7B size is friendly to fine-tuning for developers and researchers, while the 7B system is small enough to run on a phone and is built for software engineering and deep research ability.

For larger local use, IFM pointed to a dense 32B model and a sparse 36B-A4B model. Those two provide options for local hosting and on-premise servers, the institute said. Reuters placed the top of the range at a 375-billion-parameter system for enterprise deployments, while IFM identified that flagship as 375B-A23B for demanding enterprise work.

Performance claims by size class

IFM said K2 Horizon delivers top-tier performance in every size class across reasoning, mathematics, coding and agentic tasks, according to the Institute of Foundation Models press release. The institute said the 0.9B, 3.7B and 7B models set new state of the art at their respective scales.

The organization described the 0.9B model as the best-performing model of its size in math, reasoning and tool use. It described the 3.7B model as offering the best reasoning under 4B and as matching or beating many larger models.

The 7B model was described as the best-performing model under 10B, with strong software engineering and deep research ability. IFM described the dense 32B model as among the most capable dense models for local hosting and as built for heavy use on laptops and on-prem servers.

The sparse 36B model uses a Mixture of Value Attention architecture and activates only 4B parameters, which IFM said makes the model cost-effective for local hosting while outperforming many larger models. The flagship 375B model, with 23B active parameters, was built to compete with open-weight models on reasoning and agentic work and is ready for the enterprise, the institute said.

Hector Liu, director of IFM's Silicon Valley lab, said the institute was releasing an entire fleet at once rather than a single model, from one small enough to run on a watch to a flagship built for enterprise reasoning. He said every model is built to compete with the best open models at its size and ships with the weights, code, training data and methodology behind it.

Two technical ideas behind the fleet

IFM highlighted two technology innovations in the K2 Horizon announcement. The first is a "diffusion distillation" technique that generates blocks of tokens in parallel, which the institute said speeds model performance by roughly 3X without degrading response quality.

The second is a "mixture of value attention" architecture that improves reasoning without adding more computation, according to the same announcement. The 36B model implements that idea as Mixture of Value Attention, abbreviated MoVA, with 4B active parameters.

IFM said it engineered each K2 Horizon model to compete with the best open models of comparable size on reasoning, tool use and agentic tasks. It said the smallest model is designed for devices where memory, power consumption and local processing matter, while the largest model is intended for demanding reasoning and enterprise AI workloads. Developers can prototype on smaller models and scale to larger models without leaving the fleet and without changes to deployment workflows, the institute said.

How developers can get the models

Access began immediately through Hugging Face, as well as vLLM and SGLang, according to the IFM announcement. The application programming interface is available through inference partners, including Compass, Cerebras and Nebius, the institute said.

The models and code are released under the Apache 2.0 license, IFM said. That licensing detail, combined with data and recipe disclosure, is central to the institute's claim that outside groups can adapt the systems directly.

Silicon UK, in its report titled "UAE-Based Lab Releases Fully Open Source Models | Silicon UK", summarized the practical effect as allowing researchers and developers to inspect, reproduce and adapt the models for their own work. The same report noted the range from a 900 million-parameter model for smart watches and other power-constrained devices to the largest system with 375 billion parameters.

What this signals for Abu Dhabi's AI push

Reuters reported that the launch forms part of the UAE's push to establish itself as a global AI hub. IFM itself described an expansion as a global frontier lab.

IFM was launched by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in May 2025, according to the institute's announcement. It now employs AI researchers, scientists and developers across three offices in Silicon Valley, Paris and Abu Dhabi.

Xing is founder of IFM and president and university professor of MBZUAI, the institute said. In the IFM statement, Xing said open source is much more than open weights and said science works when others can see the data, follow the method, reproduce the result and improve on it.

The institute also described its wider portfolio. IFM said its model portfolio includes the K2 series as the fully open-source fleet, Jais as an Arabic large language model, and PAN as a next-generation world model for embodied reasoning and real-world simulation.