1.08B, 5.06B, 9.00B, and 34.78B stored parameters for the dense models marketed as 0.9B, 3.7B, 7B, and 32B; 37.44B / 4B active and 379.17B / 23B active for the MoE models
131,072 tokens (0.9B); 524,288 tokens (3.7B through 375B)
Apache 2.0
A same-day family of six open-weight text-generation foundation models spanning edge deployment through frontier-scale agentic workloads. The family contains dense 0.9B, 3.7B, 7B, and currently released 32B Stage1 models, a 36B mixture-of-experts model with 4B active parameters and Mixture of Value Attention (MoVA), and a flagship 375B mixture-of-experts model with 23B active parameters. The 3.7B-through-375B variants support native 512K-token contexts, while the 0.9B model supports 128K with YaRN. IFM says the models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling, with a smaller vocabulary for the 0.9B variant. The release introduces diffusion distillation for generating blocks of tokens in parallel, with IFM claiming roughly three-times faster generation without quality loss, plus dynamic routing across model sizes. The models expose controllable reasoning modes and structured tool-call and reasoning parsers. IFM released the TxT360-v2 corpus, specialized code, mathematics, reasoning, SFT, and behavior datasets, the xLLM pretraining codebase, and the Horizon post-training codebase. The exact openness status varies by variant: final weights are available for the 375B and 36B models, the 7B and 3.7B repositories include intermediate checkpoints, and the currently downloadable 32B model is explicitly named K2-Horizon-32B-Stage1 with its final checkpoint still pending. FP8, GGUF, and selected Uno variants are also available. The flagship 375B model reports 70.2 on Terminal-Bench 2.1, 42.6 on SWE-Bench Pro, 72.8 on BrowseComp, 67.7 on MCPMark, 87.3 on GPQA Diamond, and a 1441 GDPval-AA Elo rating. The 7B model reports 70.6 on SWE-Bench Verified, 59.0 on BrowseComp, and 68.0 on AA-LCR, while the 0.9B model reports 48.5 on AIME 2026 and 79.9 on HumanEval+. All benchmark results are vendor-reported and use the evaluation settings and harnesses described in the individual model cards.