ABU DHABI, Sept. 4 — The shift in open-source AI is no longer about single models. The Institute of Foundation Models (IFM) launched K2 Horizon on Wednesday, releasing what it calls the industry’s largest fully open-source fleet of AI models — six of them, spanning from a 0.9-billion-parameter version built for constrained devices to a 375-billion-parameter flagship for enterprise use. Every model in the fleet ships with its weights, code, training data, and methodologies.
IFM argues that goes beyond typical “open weights” releases, letting researchers inspect, reproduce, and adapt the systems rather than simply run them. The lineup is designed for scale diversity.
A 0.9B model targets watches and glasses; 3.7B and 7B versions handle phones and on-device apps; a dense 32B model and a sparse 36B-A4B model suit local hosting and on-premise servers. At the top sits the 375B-A23B flagship for enterprise deployments. IFM reports the three smallest models set new state-of-the-art performance at their respective scales in reasoning, mathematics, coding, and agentic tasks.
Two technical innovations stand out. A “diffusion distillation” technique generates token blocks in parallel, speeding performance by roughly 3X without quality degradation.
A “mixture of value attention” architecture improves reasoning without added computation. IFM’s dynamic model routing directs tasks to the most cost-effective model, letting developers prototype on smaller systems and scale up without changing workflows. “Open source is much more than open weights,” said Eric Xing, Founder of IFM and President and University Professor of MBZUAI.
“Science works when others can see the data, follow the method, reproduce the result, and improve on it. K2 Horizon delivers on that need. Every model in the fleet ships with its training data, recipe and evaluations.
This is open science, and we believe it’s the best path forward for AI.” Hector Liu, director of IFM’s Silicon Valley lab, framed the release as a fleet play rather than a single drop. “Developers can prototype on the smallest model, scale to the flagship, and verify every claim we make along the way,” he said.
The launch marks IFM’s expansion into a global AI frontier lab, with researchers across offices in Silicon Valley, Paris, and Abu Dhabi. IFM was launched by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in May 2025. K2 Horizon models are available immediately through Hugging Face, vLLM, and SGLang, with an API through inference partners including Compass, Cerebras, AWS, and Nebius.
Models and code sit under the Apache 2.0 license.
