K2 Horizon 0.9B

K2 Horizon 0.9B is the smallest and most compact model in the K2 Horizon fleet, a family of open-source language models developed by the Institute of Foundation Models (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). Designed specifically for highly constrained environments such as smartwatches, augmented reality glasses, and edge devices, it prioritizes efficiency while maintaining competitive performance on core reasoning and mathematical tasks.

The model was pre-trained on approximately 20 trillion tokens, featuring a significant proportion of synthetic problem-solving trajectories and reasoning-focused data. Unlike the larger models in the Horizon family, the 0.9B variant utilizes a specialized, smaller vocabulary to optimize its performance and memory footprint on limited hardware. It supports a native context window of 128K tokens (131,072) through YaRN RoPE scaling, allowing it to process substantial document lengths relative to its parameter count.

Released under the Apache 2.0 license, K2 Horizon 0.9B is part of a "fully open" release strategy. This includes not only the weights and training code but also intermediate checkpoints, training data recipes, and fine-grained evaluation logs. Despite its small size, it reports strong results on mathematical benchmarks such as AIME 2026, positioning it as a capable reasoning engine for local, on-device applications where cloud connectivity is unavailable or undesirable.