Back to feed
News Story
TestingCatalog (X)
2 sources

Institute of Foundation Models Releases K2 Horizon Open Model Family

The Institute of Foundation Models (IFM) has launched K2 Horizon, a suite of six open-source foundation models ranging from 0.9B to 375B parameters. Licensed under Apache 2.0, the release is notable for its radical transparency, including the publication of training code, data recipes, and checkpoints, targeting diverse use cases from lightweight local development to large-scale agentic applications.

SynthePulse Insight · AI deep reading

IFM Releases K2 Horizon: Six Open-Source Models Spanning 0.9B to 375B Parameters

Version 1 · 1 source

The Institute of Foundation Models (IFM) released the K2 Horizon series on September 3, 2026, comprising six open-source foundation models ranging from 0.9B to 375B parameters. Licensed under Apache 2.0, the release also includes training code and data recipes.

  • IFM releases K2 Horizon, covering model sizes of 0.9B, 3.7B, 7B, 32B, 36B, and 375B parameters.
  • Each size is optimized for specific scenarios: 0.9B for local development, 375B for long-horizon agents.
  • All models and code are licensed under Apache 2.0, with training code, data recipes, and intermediate checkpoints made public.
  • The architecture introduces MoVA routing and a Uno diffusion adapter, claiming top-tier performance in coding and agent tasks.
Open section navigationK2 Horizon Series Release Details

K2 Horizon Series Release Details

The Institute of Foundation Models (IFM) released an open-source foundation model family named K2 Horizon on September 3, 2026. The series comprises six models with a vast parameter scale range, extending from the smallest at 0.9B to the largest at 375B. This release aims to provide a tightly integrated fleet of foundation models to address computational needs and application scenarios across different levels.

Hierarchical Positioning and Use Cases

IFM has assigned clear functional positioning to each size within the K2 Horizon series. The 0.9B model is designed for local development, facilitating rapid prototyping and validation; the 3.7B model is suitable for single-node service deployment; and the 7B model is optimized for cost-sensitive deployment scenarios. Additionally, the 32B model is positioned as the mainstay for daily heavy usage, the 36B model is used for service experiments in production environments, and the largest 375B model is specifically targeted at long-horizon agent tasks.

Technical Architecture and Performance Claims

In terms of technical implementation, K2 Horizon ships with two key architectural components: MoVA, a technology that routes experts internally within the attention mechanism; and Uno, a diffusion adapter usable for generation. IFM claims that K2 Horizon provides frontier performance in coding and agent tasks, reaching top-tier levels in every size category, with the 0.9B, 3.7B, and 7B models performing particularly outstandingly. However, specific benchmark scores are not detailed in the source material, existing only as performance claims.

Open Source License and Transparency

To promote community research and development, IFM announced that all model weights and code for the K2 Horizon series are released under the Apache 2.0 license. Beyond the models themselves, the institute has made public the underlying training code, data recipes, intermediate checkpoints, logs, and evaluation results. This high degree of transparency aims to enable developers to reproduce the research process or build upon existing work for further innovation.

Credibility boundary

This report is based on an official statement from IFM forwarded by the third-party news aggregation account TestingCatalog. While core release information (such as model names, parameter scales, and license agreements) comes from official channels, specific quantitative metrics regarding 'frontier performance' and 'top-tier performance' lack independent third-party verification and constitute source claims.

Insight takeaway

By providing full-stack open-source model options ranging from edge devices to large-scale clusters, K2 Horizon lowers the entry barrier for AI applications. Its complete disclosure of training data and code helps advance the reproducibility of the open-source ecosystem.

Sources for this version

  1. The Institute of Foundation Models has released K2 Horizon, a family of six open…TestingCatalog (X)

Primary report

TestingCatalog (X)

Primary source

Same-event coverage

Also covered by 1 sources

Kimmonismus (X)T3Institute of Foundation Models Releases K2 Horizon Suite of Six Open-Source Models