The Institute of Foundation Models (IFM), a frontier lab launched by MBZUAI, has announced the release of K2 Horizon, a fleet of six open-source AI models spanning from 0.9B to 375B parameters. Available under the Apache 2.0 license on Hugging Face, the launch includes pre-training data, intermediate checkpoints, training code, and configurations. The models feature day-zero support for vLLM, SGLang, and Ollama across NVIDIA, AMD, and Cerebras hardware.
Built on a unified core architecture, the fleet allows developers to prototype on smaller models and scale to larger deployments without changing their serving stack. The models were pre-trained on roughly 20 trillion tokens, including synthetic data and explicitly detailed problem-solving trajectories. Smaller variants demonstrated high efficiency and competitive performance, while the largest model, K2-Horizon-375B-A23B, achieved top benchmark scores in reasoning and coding tasks.
IFM also published transparent safety and benchmark audit data, detailing a 3.37-point accuracy correction after auditing for reward hacking. The release aims to provide accessible, fully open-source models with consistent tooling for enterprise and developer workflows.
Why it matters
Consistent architecture across sizes allows founders to prototype cheaply on 3.7B models before seamlessly scaling to 375B.
Full open-source release under Apache 2.0 includes training code and data, offering deep customizability for enterprise deployments.
Source: marktechpost.com
