Blab Tech
IFM Releases K2 Horizon: Six Fully Open AI Models From 0.9B to 375B Parameters
On September 3 2026, the Institute of Foundation Models (IFM) in Abu Dhabi announced K2 Horizon, a suite of six foundation models that span 0.9 billion to 375 billion parameters. IFM says the collection is the largest fully open‑source set of AI models that has been released to date.
IFM’s definition of “fully open” extends beyond downloadable weights. The organization has pledged to supply training and evaluation code, training data that can be redistributed, or detailed construction recipes when the data cannot be shared. In addition, the release includes configuration files, training logs, and intermediate checkpoints that trace the models from pre‑training through post‑training reasoning and agentic stages. The goal is to let developers inspect every step of the build, reproduce the training pipeline, and adapt the models for their own use cases.
When the models went live, all six were available on Hugging Face with final weights. However, the completeness of accompanying artifacts varied. The 3.7 B and 7 B models were launched with the full set—code, data, and checkpoints. The 0.9 B variant’s documentation notes that its training data and code will be added later. The flagship 375 B‑A23B and the sparse 36 B‑A4B models were released with only final weights; IFM said that training code, raw datasets, and checkpoints would follow in subsequent updates. The 32 B model arrived as a Stage 1 checkpoint, with the final model and remaining artifacts slated for later release.
All models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling. The 0.9 B version uses a smaller vocabulary to accommodate constrained devices such as smartwatches and smartglasses. The 3.7 B and 7 B models target phones and other on‑device applications. The dense 32 B and sparse 36 B‑A4B models are intended for local hosting and on‑premises servers, while the 375 B‑A23B model is aimed at enterprise deployments that demand high capacity.
IFM’s dynamic model‑routing technique directs tasks to the most cost‑effective model, providing a clear path from prototype to production. The organization stresses that its open‑source approach goes beyond the open‑weights trend that has dominated recent AI headlines.
The release is licensed under Apache 2.0 for code and a patchwork license for data. Weights are available on Hugging Face with support for vLLM and SGLang, and an inference API is offered through partners Compass, Cerebras and Nebius.
Some members of the AI community have voiced reservations about the level of openness. Concerns include the lack of compute details such as accelerator count, training hours, and cost, as well as incomplete synthetic data pipelines. According to a comment by Nitish Garg, founder and CEO of AI super‑app CellCog, “Reasoning traces for math were rewritten into dialogues and study guides and mixed into pre‑training rather than saved for post‑training. Compute is not disclosed anywhere.” Other users on Hacker News noted that while Chinese open‑weight models release both base and instruction versions, they typically do not publish training corpora.
IFM’s director of its Silicon Valley lab, Hector Liu, described the initiative as a “360‑degree open source” effort. He said that developers can prototype on the smallest model, scale to the flagship, and verify every claim made by the organization.
The K2 Horizon release has been covered by several industry outlets, including the IFM blog, Hugging Face, and AI research sites. Researchers are evaluating the models on reasoning, coding, mathematics and agentic tasks.
As of now, the 3.7 B and 7 B models are fully documented, while the 0.9 B, 32 B, 36 B‑A4B and 375 B‑A23B models are expected to receive additional artifacts in the coming weeks. The AI community will likely monitor the pace of these updates to assess the practical openness of the fleet.
IFM’s definition of “fully open” extends beyond downloadable weights. The organization has pledged to supply training and evaluation code, training data that can be redistributed, or detailed construction recipes when the data cannot be shared. In addition, the release includes configuration files, training logs, and intermediate checkpoints that trace the models from pre‑training through post‑training reasoning and agentic stages. The goal is to let developers inspect every step of the build, reproduce the training pipeline, and adapt the models for their own use cases.
When the models went live, all six were available on Hugging Face with final weights. However, the completeness of accompanying artifacts varied. The 3.7 B and 7 B models were launched with the full set—code, data, and checkpoints. The 0.9 B variant’s documentation notes that its training data and code will be added later. The flagship 375 B‑A23B and the sparse 36 B‑A4B models were released with only final weights; IFM said that training code, raw datasets, and checkpoints would follow in subsequent updates. The 32 B model arrived as a Stage 1 checkpoint, with the final model and remaining artifacts slated for later release.
All models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling. The 0.9 B version uses a smaller vocabulary to accommodate constrained devices such as smartwatches and smartglasses. The 3.7 B and 7 B models target phones and other on‑device applications. The dense 32 B and sparse 36 B‑A4B models are intended for local hosting and on‑premises servers, while the 375 B‑A23B model is aimed at enterprise deployments that demand high capacity.
IFM’s dynamic model‑routing technique directs tasks to the most cost‑effective model, providing a clear path from prototype to production. The organization stresses that its open‑source approach goes beyond the open‑weights trend that has dominated recent AI headlines.
The release is licensed under Apache 2.0 for code and a patchwork license for data. Weights are available on Hugging Face with support for vLLM and SGLang, and an inference API is offered through partners Compass, Cerebras and Nebius.
Some members of the AI community have voiced reservations about the level of openness. Concerns include the lack of compute details such as accelerator count, training hours, and cost, as well as incomplete synthetic data pipelines. According to a comment by Nitish Garg, founder and CEO of AI super‑app CellCog, “Reasoning traces for math were rewritten into dialogues and study guides and mixed into pre‑training rather than saved for post‑training. Compute is not disclosed anywhere.” Other users on Hacker News noted that while Chinese open‑weight models release both base and instruction versions, they typically do not publish training corpora.
IFM’s director of its Silicon Valley lab, Hector Liu, described the initiative as a “360‑degree open source” effort. He said that developers can prototype on the smallest model, scale to the flagship, and verify every claim made by the organization.
The K2 Horizon release has been covered by several industry outlets, including the IFM blog, Hugging Face, and AI research sites. Researchers are evaluating the models on reasoning, coding, mathematics and agentic tasks.
As of now, the 3.7 B and 7 B models are fully documented, while the 0.9 B, 32 B, 36 B‑A4B and 375 B‑A23B models are expected to receive additional artifacts in the coming weeks. The AI community will likely monitor the pace of these updates to assess the practical openness of the fleet.
