The Pulse
IFM Releases Six Fully Open K2 Horizon AI Models
The Institute of Foundation Models has released K2 Horizon, a six-model AI fleet ranging from 0.9 billion to 375 billion parameters. IFM says the release includes model weights, code, training data or recipes, evaluations and deployment too
AI.info Team ·
Six models span 0.9 billion to 375 billion parameters in the Institute of Foundation Models’ K2 Horizon release, which the Abu Dhabi-based lab describes as the largest fully open-source model launch in AI history.
IFM, a research lab launched by Mohamed bin Zayed University of Artificial Intelligence in May 2025, publishes the fleet for reasoning, mathematics, coding, tool use and agentic tasks. The release goes beyond model weights: IFM says developers and researchers can inspect the training code, data, methodologies, evaluations and deployment tooling behind the models.
From watches to enterprise servers
K2 Horizon includes dense models at 0.9B, 3.7B, 7B and 32B parameters, alongside sparse models at 36B parameters with 4B active and 375B parameters with 23B active. IFM positions the smallest model for constrained devices such as watches and glasses, while the 3.7B and 7B versions target phones and other on-device applications.
The 32B model is aimed at local hosting on workstations and on-premises servers. The 36B-A4B model uses what IFM calls a Mixture of Value Attention architecture, activating only 4B parameters per token. The 375B-A23B flagship is designed for larger enterprise deployments and demanding reasoning and agentic workloads.
All six models share a core architecture, training methodology, interfaces and deployment tooling. The 0.9B model uses a smaller vocabulary. IFM also says its dynamic model-routing system can direct tasks to the most cost-effective model, allowing developers to move from smaller prototypes to larger deployments without changing their workflows.
IFM makes openness part of the release
Eric Xing, founder of IFM and president and university professor at MBZUAI, frames the launch as a response to the limits of open-weight releases.
“Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it.”
Eric Xing, founder, Institute of Foundation Models, and president and university professor, MBZUAI
IFM says K2 Horizon includes final weights, training code, checkpoints and either training data or data-construction recipes for the models. The lab’s website describes the project as a “glass box” approach, with the goal of letting outside researchers reproduce, fine-tune and extend the systems.
The models and code use the Apache 2.0 license. IFM makes the models available through Hugging Face, with support for vLLM and SGLang. Developers can also access APIs through inference partners including Compass, Cerebras and Nebius.
Parallel token generation and sparse attention
IFM highlights two technical features in the K2 Horizon family. A “diffusion distillation” technique generates blocks of tokens in parallel, which the lab says can improve model speed by roughly three times without reducing response quality.
The second feature, Mixture of Value Attention, is intended to improve reasoning without increasing computation. IFM applies the architecture most visibly in the sparse 36B-A4B model, which activates a small fraction of its total parameters during each operation.
Those design choices reflect the range of hardware targeted by the fleet. The smallest systems are meant to operate where memory, power consumption and local processing are limited, while the larger sparse model uses conditional computation to reduce the cost of serving a system with hundreds of billions of total parameters.
Performance claims cover three smaller models
IFM says the 0.9B, 3.7B and 7B models set new state-of-the-art results in their respective size classes. The lab describes the 0.9B model as its strongest system for math, reasoning and tool use at that scale; the 3.7B model as the leading reasoning model under 4B parameters; and the 7B model as the best-performing model under 10B parameters.
Those statements are claims from IFM’s launch materials rather than an independent ranking. The company says every K2 Horizon model was evaluated on reasoning, coding, tool-use and agentic tasks against open models of comparable size. Its public model materials include benchmark tables and intermediate checkpoints intended to show how capabilities change during training.
The 3.7B model card lists a 512K-token context window and reports scores of 70.5 percent on HMMT February 2026, 68.6 percent on SWE-bench Verified, 65.4 percent on GPQA Diamond and 25.1 percent on Terminal-Bench 2.1. The same table shows the model leading some comparisons while trailing selected competitors on scientific question answering and function calling.
A fleet designed to keep developers inside one family
Hector Liu, director of IFM’s Silicon Valley lab, said the institute chose to release a family rather than a single flagship system.
“Rather than release a single model, we’re releasing an entire fleet at once — six models, from one small enough to run on a watch to a flagship built for enterprise reasoning.”
Hector Liu, director, Silicon Valley lab, Institute of Foundation Models
That strategy gives K2 Horizon a practical advantage for teams that need to move between local experimentation and larger production workloads. A developer can test an application on a 3.7B or 7B model, run a 32B or 36B system locally, and reserve the 375B model for cases where higher capability justifies the infrastructure.
The open-data and training-methodology claims will determine how far K2 Horizon advances reproducible AI research. The models are available now through IFM’s Hugging Face collection, alongside the supporting resources the lab published for running and evaluating them.