IFM K2 Horizon models now span six sizes from 0.9 billion to 375 billion parameters, with the Abu Dhabi research institute releasing far more than downloadable weights. The September 3 launch includes code, training data, methods and intermediate checkpoints intended to let outside researchers inspect and reproduce how the systems were built.

 

The Institute of Foundation Models, created by Mohamed bin Zayed University of Artificial Intelligence in 2025, is presenting the release as an alternative to both closed frontier systems and open-weight models that reveal little about their training. Reuters independently reported the launch and its central transparency claims.

 

Six K2 Horizon Models

 

The fleet is designed as a connected family rather than six unrelated products. IFM says the models share core architecture, vocabulary, training methods, interfaces and deployment tooling, allowing developers to test an application on a small model and move it to a larger one without rebuilding the entire workflow.

 

The lineup covers a broad range of hardware and workloads:

 

  • 0.9B for constrained devices including watches and glasses
  • 3.7B for fine-tuning, scientific coding and mathematical reasoning
  • 7B for phone-scale software engineering and research tasks
  • 32B as the most capable dense model in the family
  • 36B-A4B with roughly 4 billion parameters active per token
  • 375B-A23B with roughly 23 billion parameters active per token

 

IFM says the 0.9B, 3.7B and 7B systems set new performance marks for their size classes. Those benchmark claims come from the developer and will need broader independent testing, especially across real applications where latency, memory use and reliability can differ from published evaluations.

 

Full Training Record Released

 

The most consequential part of the release may be its documentation. For each model, IFM promises final weights, pre-training and post-training code, configurations, data or data-construction recipes, intermediate checkpoints, detailed training logs and evaluation results.

 

"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, IFM's founder and the president of MBZUAI, made that distinction in the institute's launch statement. The release uses the Apache 2.0 license and is available through Hugging Face, vLLM and SGLang, while commercial API access is being offered by inference partners including Compass, Cerebras and Nebius.

 

The level of disclosure matters because final weights show only the end state of a model. Checkpoints and logs can help researchers trace when capabilities appear, test how data choices affect behaviour and repeat experiments that would otherwise depend on a lab's unpublished process.

 

Related Coverage

 

K2 Horizon Architecture Choices

 

IFM says every model was pretrained on about 20 trillion tokens. Its technical report describes a mixture of web, code, mathematics, science, multilingual material and domain-specific sources, with roughly 10 trillion synthetic tokens and nearly 17% of the corpus devoted to problem-solving trajectories with explicit reasoning.

 

The flagship uses a mixture-of-experts design with 375 billion total parameters but activates about 23 billion for each token. That structure aims to preserve the capacity of a very large model while avoiding the cost of using every parameter on every step of inference.

 

The 36B-A4B model adds what IFM calls Mixture of Value Attention, or MoVA. Instead of applying sparse expert routing only in feed-forward layers, the design extends it into attention, producing a 36-billion-parameter system that activates about 4 billion parameters per token.

 

A second technique, diffusion distillation, generates blocks of tokens in parallel. IFM claims it can raise model speed by roughly three times without reducing response quality, though outside deployment tests will be important for measuring how the gain holds up across hardware and workloads.

 

Transparency Meets Deployment Pressure

 

K2 Horizon arrives as "open" has become a disputed label. Many prominent releases provide weights that can be downloaded and fine-tuned, but omit the training corpus, construction pipeline and intermediate states needed to reconstruct the development process.

 

A reproducible model family could become useful infrastructure for universities, regulators and companies that need to audit how a system was made. The shared design across six sizes also creates a controlled research setting for comparing capability and efficiency without switching between unrelated model families.

 

Openness does not by itself verify performance, remove data-quality concerns or make a model inexpensive to retrain. The 375B system still requires substantial computing infrastructure, and the value of the release will depend on whether independent teams can reproduce IFM's results and build useful derivatives from the disclosed materials.

 

The immediate test is adoption. If researchers use the checkpoints to explain capability changes, developers move applications across the fleet, and third parties confirm the benchmark results, K2 Horizon could shift expectations for what an open foundation-model release must include.