Most open model releases deliver a single checkpoint and a benchmark table. The Institute of Foundation Models (IFM) took a broader approach last week. IFM is the frontier lab launched by MBZUAI in May 2025, and its latest release, K2 Horizon, is a fleet of six models ranging from 0.9B to 375B-A23B parameters. Alongside these models, IFM has also open-sourced the pre-training corpus, intermediate checkpoints, training code, configurations, and fine-grained logs—what the organization calls the largest fully open-source model launch in AI history.
Are the models deployable?
Yes, all six model sizes are available on Hugging Face under the Apache 2.0 license, with FP8 and GGUF builds included. Day-zero support covers vLLM, SGLang, and Ollama, and the models run on NVIDIA, AMD, and Cerebras hardware. Hosted APIs are accessible through Compass, Cerebras, and Nebius via platform.ifm.ai.
What Actually Shipped
The six models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling—with the 0.9B edition using a smaller vocabulary to accommodate its scale. This consistency is intentional: teams can prototype on the 3.7B model and scale up to 375B-A23B without changing their serving stack.
Each model was pre-trained on roughly 20 trillion tokens, ensuring a robust foundation across tasks and domains.
via MarkTechPost