K2 Horizon just shipped as six new fully open models — developers aren’t fully convinced

The Institute of Foundation Models (IFM) unveiled K2 Horizon, a group of six cutting-edge AI foundation models last week. These models, ranging from 0.9 billion to a massive 375 billion parameters, are being touted as the "largest fully open-source fleet" of AI models ever introduced. IFM goes beyond merely providing downloadable model weights; it is committed to releasing training and evaluation code, training data (where permissible), detailed construction recipes, configurations, logs, and intermediate checkpoints covering the entire AI development process.

This open-source approach allows developers to scrutinize the models' construction, reproduce their creation, and adapt them to suit their specific needs. However, it's essential to note that not all aspects of these models are immediately available at launch. While all six models come with downloadable weights, the documentation for the 0.9B, 32B, and flagship 375B models states that some training data, code, or checkpoints will be released later.

IFM founder Eric Xing emphasizes that true openness extends beyond just open weights, advocating for complete transparency where data, methods, and results are accessible for others to review, replicate, and enhance. IFM claims that K2 Horizon embodies this holistic openness, releasing every component of the model's training lifecycle—from pretraining to reasoning and agentic post-training.

For each model, IFM is releasing or has committed to release intermediate checkpoints, training data, or construction recipes. This includes architecture details, mixture compositions, training code, configurations, fine-grained logs, evaluation results, and final weights. The K2 Horizon team asserts that every model in this diverse fleet performs admirably across various tasks, including reasoning, mathematics, coding, and agentic applications.

The smallest 0.9B model is tailored for resource-constrained environments like smartwatches and smartglasses, while the larger 3.7B and 7B models offer advanced capabilities for smartphone and on-device applications. The dense 32B and sparse 36B-A4B models excel in local hosting and on-premises servers, while the powerful 375B-A23B model caters to enterprise-level demands.

Despite sharing a core architecture, vocabulary, training methodology, interfaces, and deployment tooling, the six models differ in size and performance, with the 0.9B model featuring a smaller vocabulary. IFM's dynamic model routing system ensures tasks are allocated to the most cost-efficient model, providing a seamless path from prototyping to production.

The team firmly believes that the comprehensive openness of K2 Horizon represents a significant stride in transparency, surpassing the current open-weights trend dominating AI headlines. However, questions arise about the extent of their openness. For instance, Nitish Garg, founder and CEO of AI super-app company CellCog, points out that the lack of detailed compute information and fine-grained training logs compromises the reproducibility of AI engineers' work.

While IFM shares methodologies and model training processes, the absence of complete transparency, such as accelerator counts, training hours, cost, and specific filtering heuristics for synthetic data generation, may hinder true end-to-end reproducibility. Developers may find that without these details, achieving true end-to-end reproducibility remains a challenging feat.