IFM Releases K2-Horizon-MoVA-36B-A4B GGUF Model with 512K Context Window
IFM has released GGUF quantizations for K2-Horizon-MoVA-36B-A4B, a sparse Mixture-of-Experts (MoE) model featuring 36B total parameters with only 4B active parameters per token. The model supports a native 512K context window and was launched alongside smaller GGUF variants ranging from 0.9B to 32B parameters. By achieving competitive reasoning performance while activating only 4B parameters, the model provides high computational efficiency for running large context models locally. Additionally, IFM plans to release training data, code, and intermediate checkpoints, supporting open AI research and reproducibility. The architecture incorporates Mixture-of-Values Attention (MoVA) and supports a native context length of 524,288 tokens established during mid-training. The GGUF format enables efficient deployment and inference on consumer hardware via runtimes such as llama.cpp.
## BACKGROUND
Mixture-of-Experts (MoE) is a model design that routes input tokens to specific expert sub-networks, allowing high total parameter capacity while keeping active computation low per token. GGUF is a standardized binary file format created for llama.cpp that enables fast, memory-efficient inference of quantized LLMs on CPUs and GPUs.