Abu Dhabi-based Institute of Foundation Models (IFM) unveiled K2 Horizon last week, positioning it as the "largest fully open-source fleet of AI models" to date. The lineup spans six foundation models with parameter counts between 0.9 billion and 375 billion, each accompanied by training code, evaluation methodology, and data artifacts—or detailed documentation explaining why certain data cannot be redistributed.

IFM's definition of "fully open" extends beyond downloadable weights. The organization committed to publishing training and evaluation code, training datasets where legal redistribution permits, detailed construction recipes where it does not, plus configurations, logs, and intermediate checkpoints covering the entire development pipeline from pretraining through post-training stages. The goal is enabling developers to understand model construction, replicate development processes, and customize models for specific applications.

However, this commitment did not translate to complete availability on day one. All six models shipped with downloadable weights, but documentation for the 0.9B, 32B, and flagship 375B variants indicated that certain training data, code, or checkpoints would arrive in subsequent releases. The 32B model arrived as only a Stage 1 checkpoint, with the final version 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

Comprehensive artifacts across the training lifecycle

IFM stated it is opening the complete training lifecycle for every K2 Horizon model, spanning pretraining through reasoning and agentic post-training phases. For each model, the organization is releasing or has pledged to release intermediate checkpoints, training data, or detailed data-construction recipes. Intermediate checkpoints function as snapshots captured throughout training, enabling researchers to observe model development progression and reproduce or resume specific training stages.

The broader collection of released artifacts encompasses architecture specifications, mixture compositions, training code, configurations, detailed logs, evaluation results, and final model weights. According to IFM's assessment, all model sizes demonstrate strong performance across reasoning, mathematics, coding, and agentic tasks.

The 0.9B variant targets highly constrained environments like smartwatches and smartglasses. The 3.7B and 7B models bring advanced functionality to mobile and on-device deployments. The dense 32B model and sparse 36B-A4B variant provide stronger performance for local infrastructure and on-premises server deployments. The 375B-A23B model delivers the fleet's most capable performance for demanding enterprise applications.

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

Transparency claims face developer scrutiny

While IFM clearly pursues differentiation through comprehensive openness, questions persist about whether the initiative could have been more transparent—and whether it will need to be in the future.

Openness levels diverged across models at launch. The 3.7B and 7B models arrived with the complete artifact set including weights, recipes, training code, and data. The 0.9B model's documentation indicated its data and code were forthcoming. The flagship 375B-A23B and sparse 36B-A4B initially shipped with final weights, with full training code, raw datasets, and intermediate checkpoints promised in future updates. The 32B model launched as an incomplete Stage 1 checkpoint, with the final model and remaining artifacts to follow.

The deeper question of openness authenticity hinges 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.

Nitish Garg, founder and CEO of AI super-app company CellCog, analyzed K2 Horizon's training processes and identified specific gaps. He noted that reasoning traces for mathematics were transformed into dialogues and study guides, then merged into pretraining rather than preserved for post-training stages.

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 CellCog

Publishing synthetic datasets represents progress, but if open frontier model companies release such datasets without providing complete generator prompts, seed code, or exact filtering heuristics for removing low-quality synthetic data, developers will struggle to achieve true end-to-end reproducibility.

Beyond methodologies, developers will demand execution specifics, including precise hardware topology details used to run complex models. Engineers may also seek distributed communication configurations to understand how parallel processors exchange data during training. Additionally, optimizer state records—parameters tracking ongoing model optimization for each training iteration—could provide valuable insights.

Developer discussions on this topic have been candid. The conversation has shifted somewhat from K2 Horizon specifically to Chinese laboratories emerging as prominent open-weight model suppliers, 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," 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 offered a nuanced perspective, suggesting 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 positions itself as the standard for comprehensive openness

In a recent video interview, Hector Liu, director of IFM's Silicon Valley lab, articulated the organization's positioning. "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 variant, and verify every claim IFM makes throughout the process. The observation that not everyone fully grasps 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 support for vLLM and SGLang. The K2 Horizon API is accessible through IFM's inference partners, including Compass, Cerebras, and Nebius. All K2 Horizon models and code are distributed under the Apache 2.0 license.

Source: The New Stack