LocalLLaMA Community Evaluates IFM's New K2-Horizon-MoVA-36B-A4B Model
Members of the LocalLLaMA community are seeking real-world feedback and performance comparisons for IFM's newly released K2-Horizon-MoVA-36B-A4B open-weights model. The model features a hybrid architecture combining sparse Mixture-of-Experts (MoE) feed-forward layers with a novel Mixture-of-Vector-Attention (MoVA) mechanism. As open-weights AI moves toward hybrid MoE and sparse attention designs, independent user verification is essential to confirm if high benchmark scores translate into practical coding quality. Evaluating models that activate only ~4B parameters out of ~37B helps local developers run high-capability models efficiently on consumer hardware. The K2-Horizon-MoVA-36B-A4B model stores roughly 37.4 billion parameters but only activates around 4 to 6 billion parameters per token. Users are evaluating its performance on terminal benchmarks and comparing its capabilities against established models like the Qwen series.
## BACKGROUND
Mixture-of-Experts (MoE) is an architecture that routes inputs to specialized sub-networks, enabling high model capacity while keeping active computation per token low. Terminal-Bench is a standardized benchmark designed to evaluate how autonomously AI models can perform complex computer and command-line operations.