A new standard for what "open" means in AI was set on Thursday. The Institute of Foundation Models (IFM), an Abu Dhabi-based AI institute, released K2 Horizon, a connected fleet of six language models that ships with everything from final weights to training data, intermediate checkpoints, training code and fine-grained logs — a package Reuters described as fully open-source models released with their training data and code.
Most open-weight releases still stop at the weights, leaving outsiders to guess at how the models were actually built. K2 Horizon takes the opposite approach, and its arrival is a significant moment for anyone tracking open-weight AI models across the industry.
Six models, from a watch to the enterprise
The fleet spans six sizes: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B parameters. According to IFM's announcement, the models are designed as one connected system rather than six unrelated releases.
The smallest model targets heavily constrained environments such as watches and glasses, while the 3.7B and 7B models are positioned for phones and other on-device applications. The dense 32B model and the sparse 36B-A4B model are aimed at local workstations and efficient serving, and the flagship 375B-A23B carries the fleet's strongest capabilities into demanding enterprise deployments. All six models ship with quantization support.
The models share core architecture, vocabulary, training methodology, interfaces, evaluation infrastructure and deployment tooling, with the 0.9B model using a smaller vocabulary. IFM says that consistency makes it easier to move between sizes, route work dynamically across the fleet, and study how capabilities scale.
What "fully open" actually covers
This is where K2 Horizon departs from the industry norm. For every model in the fleet, IFM says it is opening the training lifecycle from pretraining through reasoning and agentic post-training.
The release includes intermediate checkpoints, training data or detailed data-construction recipes, open architecture documentation, mixture compositions, training code, configurations, fine-grained training logs, evaluation results, and final weights. The models and code are released under the Apache 2.0 license, one of the most permissive in commercial use. Datasets are released under their applicable licenses, such as ODC-BY, and IFM commits to disclosing how data was constructed and mixed in cases where redistribution is not possible.
IFM also frames the release as the first open model family to expose the complete development process through agentic post-training. In practice, that means researchers can study how reasoning, tool use, planning and agentic behaviors emerge during training, reproduce the methods, and adapt them to new tools and domains — work that is normally impossible when only final weights are published.
The performance claims
IFM says the fleet's smaller models set new state-of-the-art results at their respective scales. The 0.9B, 3.7B and 7B models reportedly lead their size classes across mathematics, reasoning, general capability, coding and agentic tasks.
Some specifics from the announcement: K2 Horizon 0.9B scores above 48 on AIME 2026, a competition mathematics benchmark that remains brutal for models that small. The 3.7B and 7B models are highlighted for strong SWE-bench and BrowseComp results, pointing at software-engineering and multi-step browsing capability unusual for their size. The 36B-A4B model introduces a new Mixture-of-Value-Attention (MoVA) mechanism, which IFM credits for capability that outpaces what its active parameter count would suggest. The 32B and 375B-A23B models are described as ranking among the top models in their respective classes.
As with any lab-published evaluation, these figures are self-reported and independent verification will take time. But if the numbers hold up, the entry-level and mid-size tiers of the fleet could reshape what developers expect from models that run on consumer hardware.
Why Abu Dhabi keeps showing up in open AI
The release strengthens the United Arab Emirates' growing footprint in open model development. Abu Dhabi's AI ecosystem — anchored by institutions including MBZUAI, which launched the K2 Think reasoning system with G42 in 2025 — has consistently positioned open publication as a strategic differentiator, a pattern Reuters captured in its coverage of Thursday's release.
The timing also lands in a shifting market for open models. Nvidia's agreement to acquire Hugging Face for roughly $12.9 billion, announced this week, underscored just how much commercial value has accumulated around open model distribution platforms. Hyperscalers have meanwhile lined up to license open-weight models from Chinese labs, and Western AI companies published a letter warning against premature restrictions on open weights.
Against that backdrop, a frontier-adjacent fleet that publishes its entire training pipeline gives researchers, startups and governments an alternative to depending on closed APIs from a handful of US providers.
What to watch next
Three questions will determine whether K2 Horizon matters beyond the headlines. First, can the benchmark results be reproduced by independent evaluators, particularly for the tiny models where claims are most striking? Second, how permissive are the dataset licenses in practice, given that IFM acknowledges some training data cannot be redistributed and is documented by recipe instead? Third, will enterprises actually deploy a 375B sparse model outside a major cloud, or will the practical impact concentrate in the 7B-and-under tier that runs on phones and laptops?
The most likely immediate effect is on research transparency. With checkpoints, logs and data recipes public, K2 Horizon gives the field a complete worked example of how a modern reasoning and agent-capable model family is built — something the community has never had at this scale.
Stay Ahead of Every Major AI Development
Model releases are coming faster every month. Keep up with every launch and benchmark before it matters.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
