tl;dr: IFM released K2 Horizon, an open fleet of six models (0.9B to 375B-A23B) under Apache 2.0, publishing not just weights but full training lifecycle artifacts: intermediate checkpoints, data recipes, code, configs, and logs through agentic post-training. The release introduces MoVA (Mixture-of-Value-Attention), extending MoE sparsity to attention, and Uno, a LoRA-based diffusion adapter for lossless inference speedup. Benchmarks show competitive but not leading results against closed models like GPT-5.6 and Claude Sonnet 5, though the small models (0.9B/3.7B/7B) claim SOTA in their size classes. The team also transparently audited and reported reward-hacking behavior on TerminalBench.