IFM's K2 Horizon Models Promise Full Transparency, But Developers Say Key Details Are Still Missing

Abu Dhabi-based Institute of Foundation Models (IFM) unveiled K2 Horizon last week, a collection of six foundation models spanning 0.9 billion to 375 billion parameters. The company describes the lineup as the "largest fully open-source fleet of AI models" to date, with a commitment extending beyond downloadable weights to include training code, evaluation methodology, training data or detailed construction recipes, configurations, logs, and intermediate checkpoints across the entire development pipeline from pretraining through agentic post-training.

IFM's definition of "fully open" encompasses more than model weights alone. The organization aims to enable developers to examine how models were constructed, replicate their development process, and customize them for specific applications. However, this commitment did not translate into complete availability on launch day. While all six models shipped with downloadable weights, the model cards for the 0.9B, 32B, and flagship 375B versions indicated that certain training data, code, or checkpoints would be published later. The 32B release arrived only as a Stage 1 checkpoint, with the complete model still pending.

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.

Eric Xing, IFM founder and university professor at the Mohamed bin Zayed University of Artificial Intelligence

Components Released Across the Training Lifecycle

IFM states it is opening the complete training lifecycle for each K2 Horizon model, spanning pretraining through reasoning and agentic post-training phases. For every model, the organization is releasing or has pledged to release intermediate checkpoints, training data, or detailed recipes for data construction. These checkpoints represent snapshots captured throughout training, enabling researchers to study model evolution and reproduce or continue specific training stages. The broader collection of artifacts encompasses architecture specifications, mixture compositions, training code, configurations, granular logs, evaluation findings, and final weights.

Across reasoning, mathematics, coding, and agentic tasks, IFM reports that all model sizes demonstrate strong performance. The 0.9B variant targets highly constrained environments like smartwatches and smartglasses. The 3.7B and 7B models bring sophisticated capabilities to phones and on-device applications. The dense 32B model and sparse 36B-A4B model provide stronger performance for local hosting and on-premises infrastructure. The 375B-A23B model delivers the fleet's most advanced capabilities for demanding enterprise use cases.

All six models share a unified core architecture, vocabulary, training methodology, interfaces, and deployment tooling, with the exception that the 0.9B model employs a smaller vocabulary. IFM's dynamic model routing technique directs tasks to the most economical model and provides developers with a progression pathway from prototype to production. The organization emphasizes that the fully open code, training data, and recipes represent "a significant step forward in transparency" that extend "well beyond" the open-weights discussion that has dominated AI industry discourse this year.

Gaps in the Openness Promise

While IFM clearly seeks differentiation through comprehensive openness, questions remain about whether the commitment extends far enough—and whether it will need to expand further. The degree of openness diverged across models at launch. The 3.7B and 7B models arrived with the complete artifact collection, including weights, recipes, training code, and data. The 0.9B model card indicated its data and code were still forthcoming. The flagship 375B-A23B and sparse 36B-A4B models initially shipped with final weights, with full training code, raw datasets, and intermediate checkpoints promised in subsequent releases. The 32B model debuted as an incomplete Stage 1 checkpoint, with the final version and remaining artifacts to follow.

The deeper question of openness authenticity centers on granular aspects of model training—details that will either satisfy or disappoint developers. If published data for each model size and synthetic generation pipelines lack full reproducibility, AI engineers are unlikely to be impressed.

Reasoning traces for math were rewritten into dialogues and study guides and mixed into pretraining rather than saved for post-training. 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.

Nitish Garg, founder & CEO of AI super-app company CellCog

Releasing synthetic datasets represents progress, but if open frontier model companies do so without providing complete generator prompts (text inputs directing AI models to create synthetic training data), seed code (the core code controlling and launching dataset generation), or precise filtering heuristics (methods for removing low-quality synthetic data), developers will struggle to achieve true end-to-end reproducibility.

Beyond methodologies, developers will seek execution specifics, including exact hardware topology details for running complex models. Engineers may also want distributed communication configurations to examine how parallel processors exchange data during training. Optimizer state records—parameters tracking ongoing model optimization for each training iteration—represent another potential gap.

Developers discussing this topic have been candid. The conversation has shifted from K2 Horizon to Chinese labs emerging as prominent suppliers of open-weight models, though their training data and full training stacks typically remain proprietary.

When one user claimed that "Chinese models these days don't even release pre-trained weights anymore" and developers only receive finished post-trained products, user culi countered, "No? That's absolutely not true. Qwen, GLM, Kimi, DeepSeek, etc all consistently release both the post-trained 'Instruct/Chat' versions and the underlying 'base' (pre-trained) weights."

Hacker News user thepasch suggested that genuine open-weight openness exists but has boundaries. "Inference code, yes, but 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."

IFM's Vision of Complete Openness

In a recent video interview, Hector Liu, director of IFM's Silicon Valley lab, stated, "In AI, recently, people have some confusion about the [term] open source. People sometimes open-weight their final model, but they don't let you know how things are trained, how production is done… so at IFM we are the pioneer of 360 [degree] open source or fully open source."

Liu added that developers can prototype using the smallest model, scale to the flagship, and verify each claim IFM makes throughout the process. The observation that not everyone fully grasps the distinctions between different open approaches to technology is hardly surprising, yet the perception gap here is evident.

Model weights are available on Hugging Face, with launch-day compatibility for vLLM and SGLang. The K2 Horizon API is accessible through IFM's inference partners, including Compass, Cerebras, and Nebius. K2 Horizon models and code are distributed under the Apache 2.0 license.

Source: The New Stack