The Institute of Foundation Models (IFM) has released K2 Horizon, a family of six AI models ranging from 0.9 billion to 375 billion parameters. The unusual part is not only the range of model sizes: IFM is publishing a substantially larger portion of the development stack than a typical open-weight release.
The six variants are 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B. IFM says they share a common architecture, interfaces and deployment tooling so developers can move between smaller local models and larger systems without rebuilding the surrounding workflow. The smallest models target constrained and on-device use, while the 32B and 36B variants are positioned for local or on-premise hosting and the 375B-A23B model for heavier enterprise workloads.
More than an open-weight release
IFM is releasing final weights and code under Apache 2.0, along with intermediate checkpoints, training configurations, evaluation results, logs and training data where licensing permits redistribution. Where data cannot be redistributed, the institute says it provides construction methods and mixture recipes instead.
That distinction matters because many models described as open source expose downloadable weights but not the information needed to reproduce how the model was trained. Reuters independently reported that K2 Horizon includes weights, training data, code, methodologies and intermediate checkpoints, framing the release as a challenge to increasingly closed model-development practices.
One family across very different hardware budgets
The product proposition is a model fleet rather than a single flagship. IFM positions the 0.9B model for highly constrained devices, the 3.7B and 7B variants for mobile and developer use, the 32B and sparse 36B-A4B models for local hosting, and the 375B-A23B model for demanding reasoning and agentic workloads.
Day-one distribution includes Hugging Face plus support for vLLM and SGLang, while IFM says API access is available through inference partners including Compass, Cerebras, AWS and Nebius. This gives developers both self-hosted and hosted paths rather than forcing a single deployment model.
Performance claims still need independent testing
IFM reports top-tier results across reasoning, mathematics, coding and agentic tasks, including state-of-the-art claims for several smaller sizes. Those are vendor claims and should not be treated as settled benchmark leadership until independent evaluation catches up. Artificial Analysis already lists the 375B-A23B model and reports an Intelligence Index score of 47, providing an early external reference point, but that does not validate every benchmark claim made for the full six-model family.
The durable significance of K2 Horizon is therefore its combination of model breadth and unusually deep release transparency. For teams that care about reproducibility, local deployment or inspecting more than a final checkpoint, it offers a materially different proposition from both closed APIs and conventional open-weight releases.