K2 Horizon 32B
K2 Horizon 32B is a large dense decoder-only language model developed by the Institute of Foundation Models (IFM) as part of the K2 Horizon model family. Released under the Apache 2.0 license, the model features approximately 34.8 billion stored parameters (including its embeddings) and supports an extensive 524,288-token context window carried forward from its midtraining stages. Designed to handle demanding reasoning, coding, and multi-step tasks, the model integrates a native reasoning architecture where thoughts are processed and separated prior to generating the final output.
The model architecture comprises 64 transformer layers with a hidden size of 5,120 and utilizes a vocabulary size of 250,624 tokens. It is tailored for high-capacity local deployments and workstation environments, positioned as a dense counterpart to sparse mixture-of-experts options within the same family. Official inference integrations support tool-calling and reasoning parsers optimized for frameworks such as vLLM and SGLang.
Create with Crafiq
Generate images, 3D models, video and audio in one studio.
Explore the studioHow K2 Horizon 32B ranks
K2 Horizon 32B is highlighted in the table below. Switch the metric to see how the ordering changes.