IFM's K2-Horizon MoVA Ships as a 36B MoE Model

IFM has published K2-Horizon-MoVA-36B-A4B, an open-weight large language model built on a mixture-of-experts (MoE) design. The model carries 36 billion total parameters but activates only around 4 billion for any given token, according to its Hugging Face repository. It is the first entry in the K2-Horizon family and focuses on text generation.

The sparse MoE approach is the model's defining feature. Rather than firing every parameter on each forward pass, the network routes tokens to a small subset of specialized experts. That lets a model hold a large parameter budget while keeping the per-token compute cost closer to a much smaller dense model — a trade-off that has become increasingly common as teams try to balance quality against inference expense.

Why it matters

MoE architectures have moved from research curiosity to a practical deployment strategy, and a 36B/4B-active configuration lands in a range that could be attractive for teams with modest hardware.

• Total parameters: 36B, distributed across experts
• Active parameters: roughly 4B per token
• Modality: text-only language generation
• Weights: openly available under a custom license

Several details remain unstated in the initial listing, including context length and the specific license terms, which IFM marks simply as "other." As with any new open release, real-world evaluation will determine how the efficiency claims translate into usable performance. For now, the weights are available for developers to download and test directly from the model page.