IFM’s K2 Horizon Release: Six AI Models With Data, Code and Training Records

IFM, the Institute of Foundation Models affiliated with Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence, introduced K2 Horizon on September 3, 2026. The release is a connected family of six AI foundation models—0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B parameters—released with models and code under the Apache 2.0 license. 29

The central claim is not simply that the models can be downloaded. IFM says K2 Horizon opens much more of the development record than a typical open-weight release, with the goal of making model-building work inspectable and reproducible.

What K2 Horizon includes

For each model, IFM says it is releasing final weights alongside material covering the lifecycle from pretraining through reasoning and agentic post-training. The published material includes:

• training data or detailed data-construction recipes;
• open architecture and mixture compositions;
• training code and configurations;
• intermediate checkpoints and fine-grained training logs;
• evaluation results; and
• final model weights. 9

Reuters similarly reported that the package includes weights, training data, code, methodologies and intermediate checkpoints, enabling researchers to retrace development and reproduce results. 2

There is an important qualification: IFM describes providing training data or recipes, rather than necessarily redistributing every dataset in full. That distinction matters where data licensing or redistribution limits apply. 9

Why this goes beyond “open weights”

An open-weight model gives developers the learned parameters needed to run or adapt a system. But weights alone do not reveal the complete path that produced them: the training-data composition, preprocessing decisions, code, training configuration, intermediate model states and post-training procedures may remain unavailable.

K2 Horizon’s stated approach is to expose those additional layers. In practice, that can let independent teams check how a result was produced, attempt to replicate stages of training, test technical claims, and adapt the work for other research or deployment needs. 29

That is the meaningful distinction behind IFM founder Eric Xing’s formulation that openness should let others “see the data, follow the method, reproduce the result, and improve on it.” 1

How it compares with open-weight and closed-model approaches

The comparison is best understood as a spectrum of disclosure rather than a label alone.

• Open weights: A developer may publish a model for download while disclosing limited information about its provenance or full training process. Reuters reported that IFM’s release was intended as a contrast to an industry trend toward more secretive AI development. 2
• K2 Horizon’s fully open approach: IFM says it is publishing not only weights, but also data or construction recipes, code, configurations, checkpoints, logs and evaluations. 9
• Closed service access: Leading proprietary models are commonly delivered through hosted products or APIs rather than with a complete public development lifecycle. The sources provided here do not substantiate a detailed company-by-company comparison, so claims about specific Chinese developers, OpenAI or Anthropic should be treated cautiously.

The practical difference is verification. A user can evaluate an open-weight model’s behavior, but a fuller release gives researchers substantially more evidence to examine the process behind that behavior.

Reproducibility, oversight and competitive performance

Xing told Reuters that IFM wanted to establish a reference point for what a genuinely open model release could look like. 2 The stated aim is reproducible science rather than trust in an opaque final checkpoint.

That transparency also has potential relevance beyond research. More complete artifacts can give policymakers, regulators and public-interest groups material to scrutinize claims about model capabilities, methods and risks, rather than relying entirely on a developer’s summary of its work. 2

IFM also argues that openness need not mean lower capability. Its launch materials position K2 Horizon across reasoning, mathematics, coding, tool use and agentic tasks, while presenting the six systems as one connected model family. Those are IFM’s performance claims, not independent validation. 19

Six sizes for different deployment environments

K2 Horizon ranges from very small models intended for constrained hardware to a large mixture-of-experts flagship.

Model	Intended use described by IFM
0.9B	Resource-constrained edge devices, including wearables such as watches and smart glasses. 1
3.7B	On-device and phone-oriented use. 1
7B	On-device use, with an emphasis on software-engineering and deeper research tasks. 1
32B	Local hosting, including laptops, workstations and on-premise servers. 1
36B-A4B	A sparse model for local and on-premise deployments. 19
375B-A23B	The flagship for demanding reasoning, research, software-engineering and longer-horizon agentic work. 19

IFM says the common architecture, interfaces, tooling and routing approach are designed to let developers move from smaller to larger models without changing deployment workflows. 1

Why the launch matters for the UAE

The release supports the UAE’s ambition to build a larger role in frontier AI. IFM was launched by MBZUAI, and K2 Horizon gives the institution a globally usable foundation-model family whose development artifacts are intended to be accessible to researchers and developers beyond Abu Dhabi. 12

Whether the release becomes a durable standard for AI openness will depend on how usable the materials prove to be and whether outside teams can independently reproduce and extend the work. But K2 Horizon makes a clear proposition: in AI, openness should mean more than the ability to download a finished model. 29