K2 Horizon 0.9B

K2 Horizon 0.9B is the compact dense member of IFM's (MBZUAI) fully open K2 Horizon family: a 0.9B-class decoder-only reasoning model (28 layers, smaller 64K vocabulary; about 1.08B parameters in the released safetensors) pretrained on 5T tokens and post-trained with multi-teacher distillation from math/code, STEM, and instruction-following experts followed by on-policy distillation. Supports English and Chinese and a 131,072-token (128K) context window via YaRN RoPE scaling. Text in, text out, with per-request reasoning effort (high recommended) and JSON/XML tool-call formats. Released 2026-09-03 under Apache 2.0 with BF16 weights plus a GGUF build, an optional Uno discrete-diffusion LoRA adapter for faster decoding, intermediate checkpoints, and training code. No priced provider is cataloged yet.