Technology
IFM releases K2 Horizon open AI model family ranging from 0.9B to 375B parameters
The release includes open weights, training code, checkpoints, and data recipes under Apache 2.0 and open licenses.
The short version
- IFM released K2 Horizon, a suite of six AI models scaled from 0.9 billion to 375 billion parameters for edge, workstation, and enterprise use.
- The release exposes the entire training lifecycle under the Apache 2.0 license, including intermediate checkpoints, training code, configurations, logs, and data recipes.
- IFM claims the smaller models achieve state-of-the-art results in their respective size classes and introduces a sparse attention architecture called Mixture-of-Value-Attention (MoVA).
- Independent verification of the models' benchmark performance and third-party replication across various hardware environments remain to be seen.
Key facts
- IFM announced the release of K2 Horizon, a family of six AI models: 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B.[Hacker News]
- The release provides model weights, training code, intermediate checkpoints, and configurations under the Apache 2.0 license, alongside data recipes or open datasets under licenses such as ODC-BY.[Hacker News]
- Each model in the fleet was pretrained on approximately 20 trillion tokens, including roughly 10 trillion synthetic tokens and problem-solving reasoning trajectories.[Hacker News]
- The 36B-A4B model incorporates a Mixture-of-Value-Attention (MoVA) mechanism that activates approximately 4 billion parameters per token while approaching the performance of the dense 32B model, according to IFM.[Hacker News]
- The largest model in the fleet, the 375B-A23B sparse mixture-of-experts model, activates approximately 23 billion parameters per token.[Hacker News]
What remains uncertain
- Performance claims, benchmark scores across evaluations such as AIME 2026 and SWE-bench, and state-of-the-art status rely entirely on IFM's self-reported findings and have not yet been independently verified.[Hacker News]