K2 Horizon debuts six fully open AI models—yet transparency gaps spark developer skepticism

The Institute of Foundation Models (IFM), based in Abu Dhabi, unveiled K2 Horizon last week—a lineup of six foundation models ranging from 0.9 billion to 375 billion parameters. The organization positions the release as "the largest fully open-source fleet" in existence, but initial reception from the developer community suggests the concept of "fully open" needs refinement.

Unlike typical open-weight model releases that distribute only model parameters, IFM has committed to publishing training code, evaluation frameworks, training data (or detailed recipes where redistribution isn't feasible), configuration files, logs, and intermediate training checkpoints. The goal is to enable developers to examine how models were constructed, reproduce the process, and customize them for their own applications.

The staged rollout and uneven transparency#

The launch itself revealed inconsistencies in the transparency promise. While all six models arrived with downloadable weights, the documentation for the 0.9B, 32B, and flagship 375B variants indicated that training data, code, or checkpoints would come later. The 32B release came as an incomplete Stage 1 checkpoint rather than a finalized model, with the complete version still in development.

IFM founder Eric Xing framed the initiative this way: "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." The flagship 3.7B and 7B models shipped with the complete artifact set, while larger and smaller variants arrived incomplete—a pattern that undermined the unified openness narrative.

What developers actually want (and aren't getting)#

Nitish Garg, founder of AI platform CellCog, identified the core gap in his analysis: "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 reproducibility challenge runs deeper. Even when synthetic training datasets are shared, developers expect accompanying details that rarely materialize:

• Generator prompts—the exact text inputs used to instruct models to create synthetic data
• Seed code—the foundational code controlling dataset generation
• Filtering heuristics—the criteria used to remove low-quality synthetic examples