← Back to feed

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi has released K2 Horizon, a connected fleet of six foundation models under Apache 2.0 licensing. The release covers every size tier from wearable edge to enterprise cluster: 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B.

MBZUAI describes it as the most comprehensive open model release to date, and the claim is defensible. The lab is releasing not just weights but the full development pipeline: intermediate training checkpoints, training data or detailed data construction recipes, architecture specifications, mixture compositions, training code, evaluation code, configurations, and fine-grained training logs. All of that spans pretraining through reasoning and agentic post-training.

Performance Across the Fleet

The three smallest models — 0.9B, 3.7B, and 7B — achieve what IFM calls state-of-the-art results in their respective size classes across mathematics, reasoning, general capability, coding, and agentic tasks.

The 36B-A4B introduces a new technique called Mixture-of-Value-Attention (MoVA), which the lab says delivers exceptional capability per active parameter. With only 4B active parameters at inference time, the model outperforms some much larger dense models on the benchmarks IFM ran.

The 32B and 375B-A23B sit at the high end, competitive in their classes for enterprise deployment. The 375B-A23B is a MoE model with 23B active parameters at inference, comparable in inference cost to models roughly a third of its total size.

Agentic First

The most significant claim is architectural: K2 Horizon is billed as the first open model family to expose the complete development process through agentic post-training. This matters because agentic post-training — the RL and supervised stages that teach a model to use tools, plan across steps, and recover from errors — is where labs have historically been most secretive.

By releasing checkpoints, data recipes, and training logs for every stage including agentic, MBZUAI is allowing researchers to study how reasoning, tool use, and planning capabilities emerge and to reproduce the methods with different tools, environments, or domains.

Connected Fleet Design

All six models share a core architecture, vocabulary, training methodology, interfaces, evaluation infrastructure, and deployment tooling. The 0.9B uses a smaller vocabulary optimized for constrained environments; otherwise the design is consistent across all sizes.

The shared foundation means routing work dynamically between model sizes — a 0.9B handling simple queries, a 32B handling complex ones — is architecturally straightforward rather than requiring cross-family adapters.

Context

The release follows Qwen3.8’s Apache 2.0 open weights (2.4T parameters) last month and Kimi K3’s open release in July. The open frontier is getting substantially more competitive at the top end. K2 Horizon’s differentiation is not raw scale — the 375B-A23B does not break new benchmark ceilings on the open leaderboards — but the completeness of the release. Labs trying to reproduce frontier agentic training without access to OpenAI, Anthropic, or Google’s infrastructure now have a documented path through the full pipeline.

IFM is an Abu Dhabi state-backed research institute. The Apache 2.0 license imposes no commercial restrictions.