The Institute of Foundation Models (IFM) has released six open AI models under its K2 Horizon project, claiming it’s the “largest fully open-source fleet of AI models” available. This release marks a significant milestone in the AI community, as it aims to set a new standard for transparency and reproducibility in AI development.
Based in Abu Dhabi, the capital of the United Arab Emirates, IFM has positioned itself as a leader in open-source AI initiatives. The models, ranging from 0.9 billion to 375 billion parameters, are not just about downloadable weights; they encompass a full suite of resources including training code, data, configurations, and checkpoints from pretraining to post-training. This holistic approach ensures that developers have access to the entire development lifecycle of these models.
What Makes K2 Horizon Different
IFM defines “fully open” as providing everything needed to inspect, reproduce, and adapt the models. This includes training data (where redistribution is possible) or detailed recipes for its creation, along with logs and checkpoints throughout the training process. By offering such granularity, IFM aims to support a culture of open science in AI, where researchers and developers can build upon each other’s work with confidence.
The goal is to offer transparency and reproducibility, allowing developers to understand how the models were built and adapt them for their own projects. However, while all models have downloadable weights, some training data, code, and checkpoints for the larger models are still pending release. This phased approach ensures that the community can begin experimenting with the models immediately, while IFM continues to refine and release additional resources.
Historically, AI model releases often lacked this level of detail, making true reproducibility difficult. K2 Horizon’s approach, while not yet complete for all models, represents a notable shift towards greater openness in AI development. By setting a precedent for full disclosure, IFM is encouraging other organizations to follow suit, ultimately benefiting the entire AI ecosystem.
Model Details and Performance
The six models are designed for various use cases, demonstrating IFM’s commitment to versatility and accessibility. The 0.9B model is optimized for constrained devices like smartwatches, offering advanced AI capabilities in resource-limited environments. The 3.7B and 7B models are tailored for on-device applications, providing a balance between performance and efficiency for smartphones and similar devices.
Larger models like the 32B and 36B-A4B are suited for local hosting, offering robust performance for on-premises servers. The flagship 375B-A23B model is aimed at enterprise deployments, delivering state-of-the-art capabilities for large-scale, high-demand applications. Each model is designed to excel in specific domains, from reasoning and mathematics to coding and agentic tasks, ensuring broad applicability across industries.
All models share a core architecture, vocabulary, and training methodology, which simplifies integration and adaptation. IFM’s dynamic model routing technology intelligently directs tasks to the most efficient model, streamlining the development process from prototype to production. This seamless transition is a key advantage for developers looking to deploy AI solutions quickly and effectively.
Openness and Developer Concerns
While IFM emphasizes its commitment to openness, some developers remain skeptical about the completeness of the released resources. The 3.7B and 7B models shipped with complete artifacts, but the 0.9B, 375B-A23B, and 36B-A4B models are still missing key components like training data and code. This has raised concerns about the ability to fully reproduce and inspect these models.
According to Nitish Garg, founder & CEO of AI super-app company CellCog, in his analysis of K2 Horizon and its model training processes, “Compute is not disclosed anywhere: no accelerator count, no hours, no cost. For a release whose thesis is inspectability, that is the one obvious hole, and the fine-grained training logs, when they arrive, may fill it.” The narrative here suggests that releasing synthetic datasets is good.
Still, if open frontier model companies do this without also providing the full generator prompts (text inputs that direct AI models to create synthetic training data), seed code (as it sounds, core code that controls and initiates the dataset generation process), or exact filtering heuristics (where low-quality synthetic data needs to be cleaned or removed), then developers will find that true end-to-end reproducibility is hard, problematic, and in some cases impossible. These concerns highlight the need for even greater transparency in the AI development process.
The discussion has also highlighted the contrast with Chinese open-weight models, which often release only the finished post-trained product but keep training data and processes closed. This raises questions about the limits of openness in the AI community and the varying interpretations of what it means to be “open source.” IFM’s approach, while ambitious, shows the ongoing debate about the boundaries and expectations of openness in AI. IFM’s”
IFM’s Vision for Open Source AI
IFM aims to provide a complete development lifecycle, from prototyping to production, with verifiable claims at every stage. This vision is rooted in the principles of open science, where transparency and reproducibility are vital.
While K2 Horizon represents a significant step towards open AI, its success will depend on addressing developer concerns and fully delivering on its promises of transparency and reproducibility. By continuing to release additional resources and engaging with the community, IFM can solidify its position as a leader in open-source AI and inspire further innovation in the field.
