IFM Releases K2 Horizon: Six Open Models from 0.9B to 375B
Original: K2 Horizon: A connected fleet of six open models
Why This Matters
The first fully open agentic model fleet spanning edge to enterprise sets a new transparency benchmark in open-source AI development.
The Institute of Foundation Models (IFM) released K2 Horizon on September 3, 2026 — a fleet of six models (0.9B, 3.7B, 7B, 32B, 36B-A4B, 375B-A23B) under Apache 2.0, covering edge devices to enterprise, with 0.9B, 3.7B, and 7B setting new state-of-the-art benchmarks at their scales.
IFM released K2 Horizon, a connected fleet of six open models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. The release covers the full training lifecycle — pretraining through reasoning and agentic post-training — including intermediate checkpoints, training data or data-construction recipes, open architecture, mixture compositions, training code, configurations, fine-grained logs, evaluation results, and final weights. Models and code are released under Apache 2.0; datasets under applicable licenses such as ODC-BY.
The 0.9B, 3.7B, and 7B models achieve state-of-the-art results in mathematics, reasoning, coding, and agentic tasks at their respective scales. Notably, the 0.9B model scores above 48 on AIME 2026. The 36B-A4B introduces a new Mixture-of-Value-Attention (MoVA) mechanism, outperforming some larger models per active parameter. The 32B and 375B-A23B rank among the top in their classes. All six models share core architecture, vocabulary, and training methodology, with quantization support included. The 0.9B model targets wearables and glasses; 3.7B and 7B target on-device mobile use; 32B and 36B-A4B target local workstations; and 375B-A23B targets enterprise deployments. IFM describes K2 Horizon as the first fully open model family to expose the complete agentic post-training development process.