K2 Horizon 0.9B

K2 Horizon 0.9B - the compact dense member of the K2-Horizon fleet

K2 Horizon 0.9B is a 0.9B-class dense decoder-only reasoning model from MBZUAI's Institute of Foundation Models. It is the smallest model in the six-model K2-Horizon fleet and is designed for highly constrained environments such as watches and glasses. Despite its size, it carries a 128K context window and a full reasoning and tool-use interface.

What sets it apart

Seven experts, one merged model

The most distinctive part of the training pipeline is the RL stage. Instead of one generalist run, IFM trains seven expert models from the same midtraining checkpoint, each focused on a domain such as competition math, code, instruction following, or STEM. The experts are then merged into a single model. A follow-up MOPD stage uses on-policy distillation to fix the performance degradation that weight merging normally causes, aligning the specialist capabilities in the model's behavior rather than its weights.

Built for the smallest devices

Within the fleet, the 0.9B model is the edge endpoint: it shares the K2-Horizon architecture, training methodology, interfaces, and deployment tooling of its larger siblings, with a smaller vocabulary. This makes it usable on watches and glasses while keeping the same reasoning controls, tool-call parsing, and serving recipes as the rest of the fleet.