K2 Horizon MoVA 36B-A4B

K2 Horizon MoVA 36B-A4B is the sparse member of the K2 Horizon model family from the Institute of Foundation Models (IFM). It is a Mixture-of-Experts (MoE) model with Mixture-of-Values attention (MoVA): it stores 36B parameters and activates about 4B per token.

The model provides a native 524,288-token context window (512K) from the midtraining stage onward. It is text-in / text-out, supports reasoning controlled via a reasoning-effort parameter, and tool use in agentic scenarios. Weights are released under Apache 2.0, and the training data, recipe and training code are stated to be released.

It is served via vLLM and SGLang in BF16 precision, with GGUF quantizations available (llama.cpp, Ollama, LM Studio, Jan). Reported scores were obtained at high reasoning effort.