The Institute of Foundation Models (IFM) in Abu Dhabi just released K2 Horizon, a fleet of six AI models ranging from 0.9 billion to 375 billion parameters, claiming to be the largest fully open-source AI model collection ever released. But developer reaction reveals a growing tension in the AI industry: what companies call "open" often falls short of what researchers actually need to understand, reproduce, and improve upon these systems.

IFM's definition of "fully open" goes beyond simply releasing downloadable model weights, which is what most companies do. The organization committed to publishing training code, evaluation code, training data or detailed recipes for creating that data, plus configurations, logs, and intermediate checkpoints that show how models develop during training. The goal is ambitious: let developers inspect every step of how these models were built and adapt them for their own work.

Yet the rollout revealed cracks in that promise. Only two of the six models, the 3.7 billion and 7 billion parameter versions, shipped with the complete artifact set at launch. The smallest 0.9 billion parameter model had training data and code still forthcoming. The flagship 375 billion parameter model arrived with weights only, with full training code and intermediate checkpoints promised later. The 32 billion parameter model shipped as an incomplete Stage 1 checkpoint, with the final version still in development.

What's Actually Missing From K2 Horizon's "Open" Release?

The gaps go deeper than incomplete rollouts. Developers and AI engineers have identified several critical pieces of information that remain undisclosed, making true reproducibility difficult or impossible. According to Nitish Garg, founder and CEO of AI company CellCog, the most glaring omission involves compute details.

"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," stated Nitish Garg, founder and CEO of CellCog.

Nitish Garg, Founder and CEO, CellCog

Beyond compute, developers flagged several other missing components that prevent full reproducibility:

• Generator Prompts: The exact text inputs used to direct AI models to create synthetic training data were not shared, making it impossible for others to recreate the same training datasets.
• Seed Code: The core code that controls and initiates dataset generation processes remains undisclosed, limiting developers' ability to understand how training data was produced.
• Filtering Heuristics: The specific methods used to clean or remove low-quality synthetic data from training sets were not provided, obscuring a crucial step in data preparation.
• Hardware Topology Details: Precise information about the physical arrangement and configuration of computing hardware used during training was not released.
• Distributed Communication Configurations: Details about how parallel processors exchanged data during training remain hidden.

Garg also noted that reasoning traces for mathematics were rewritten into dialogues and study guides, then mixed into pretraining rather than kept separate for post-training evaluation. This architectural choice affects how the models learned and how researchers can understand that learning process.

How Does K2 Horizon Compare to Chinese Open-Weight Models?

The conversation around K2 Horizon's openness quickly shifted to a broader industry comparison. Developers on Hacker News pointed out that Chinese AI labs have actually been more consistent about releasing both pre-trained base models and fine-tuned versions, even if their training data and full training stacks remain closed.

When one developer claimed that "Chinese models these days don't even release pre-trained weights anymore," another user pushed back with specifics. Chinese labs including Qwen, GLM, Kimi, and DeepSeek consistently release both the post-trained "Instruct" or "Chat" versions and the underlying base pre-trained weights, according to developer discussion.

However, Hacker News user thepasch noted an important caveat: while Chinese models do release inference code and weights, "the specifics of their training process, as well as the training of the vast majority of all other open-weight models, are still a complete black box, and I can't think of any Chinese model that made its training corpus public".

This comparison highlights a paradox in the open-source AI landscape. Many companies, whether based in the Middle East, China, or elsewhere, release model weights but keep the training methodology proprietary. IFM's K2 Horizon attempted to go further by committing to share training code and data, but incomplete rollouts and missing compute details suggest the industry still has not solved the reproducibility problem.

What Does "Open Source" Actually Mean in AI?

Eric Xing, founder of IFM and a university professor at the Mohamed bin Zayed University of Artificial Intelligence, articulated the organization's philosophy on what true openness should look like.

"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. K2 Horizon delivers on that need. Every model in the fleet ships with its training data, recipe, and evaluations. This is open science, and we believe it's the best path forward for AI," explained Eric Xing, founder of IFM.

Eric Xing, Founder, Institute of Foundation Models

Xing's statement captures the tension at the heart of this story. The AI industry has conflated "open weights" with "open source," but they are not the same thing. Open weights means the model parameters are downloadable. Open source, in the traditional software sense, means the entire development process, including source code, data, and methodology, is transparent and reproducible.

Hector Liu, director of IFM's Silicon Valley lab, framed the distinction differently in a recent video interview. He said that many companies "open-weight their final model, but they don't let you know how things are trained, how production is done." IFM positioned itself as "the pioneer of 360-degree open source or fully open source," suggesting a commitment to transparency across all dimensions of model development.

Yet the actual K2 Horizon release suggests that achieving true 360-degree openness remains difficult. The incomplete rollouts, missing compute details, and undisclosed training methodologies indicate that even well-intentioned efforts to open-source AI models still fall short of the scientific reproducibility standard that Xing described.

Steps to Evaluate Open-Source AI Model Claims

For developers and organizations considering open-source AI models, the K2 Horizon experience offers practical lessons about what to look for beyond marketing claims:

• Verify Complete Artifact Release: Check whether all six components are available at launch: model weights, training code, training data or detailed data recipes, configurations, logs, and intermediate checkpoints. Incomplete releases mean reproducibility will be limited.
• Demand Compute Transparency: Ask for specific details about hardware used, including accelerator count, training hours, and estimated cost. Without this information, you cannot assess whether you have the resources to reproduce or adapt the model.
• Inspect Synthetic Data Methods: If the model was trained on synthetic data, request the generator prompts, seed code, and filtering heuristics. Without these, you cannot recreate the training dataset.
• Review Training Methodology Documentation: Ensure the release includes distributed communication configurations, hardware topology details, and optimizer state records. These details are essential for understanding how the model was actually built.
• Compare Against Industry Standards: Benchmark the release against both traditional open-source software standards and what competing AI labs have disclosed. This helps you assess whether the "open" claim is genuine or marketing.

The K2 Horizon release represents genuine progress toward more transparent AI development. The commitment to eventually release training code, data, and intermediate checkpoints goes beyond what most companies offer. However, the incomplete rollout and missing compute details demonstrate that the industry still has not fully solved the reproducibility challenge.

As AI models become more central to critical applications, the pressure for true transparency will likely increase. Developers and researchers will continue to demand not just downloadable weights, but the full training methodology, data, and computational details needed to understand, reproduce, and improve upon these systems. K2 Horizon's mixed reception suggests that companies claiming to be "fully open" will face increasingly rigorous scrutiny from the developer community.