K2 Horizon: A connected fleet of six open models
Points and comments are a snapshot, not live.
IFM releases K2 Horizon, a fully open fleet of six AI models spanning 0.9B to 375B parameters.
IFM released K2 Horizon, a family of six models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B, with the three smallest setting state-of-the-art results in their size classes. The release includes intermediate checkpoints, training code, data recipes, logs, and evaluation results under Apache 2.0. The 36B-A4B model uses a new Mixture-of-Value-Attention (MoVA) mechanism. All models share core architecture and training methodology. Pretraining used ~20 trillion tokens, with 17% being reasoning trajectories. Post-training includes branches for reasoning, coding, and agentic tasks.
What commenters are saying
Some commenters expressed model fatigue, comparing the pace to early PC CPUs or smartphones, suggesting most users only need 'good enough' models. Others emphasized the importance of fully open models for transparency and societal trust, though they acknowledged this is not yet mainstream. The thread also noted a login wall on the blog and pointed to the respective Hugging Face collection for charts and details.