IFM releases K2-Horizon, a 375B open-weight MoE

IFM has published K2-Horizon-375B-A23B, an open-weight large language model built around a mixture-of-experts (MoE) architecture. The model spans 375 billion total parameters but activates roughly 23 billion for any given token, a design meant to deliver the capacity of a very large model without paying the full compute bill on every forward pass. The weights are available on Hugging Face.

The release is positioned as a flagship, and IFM lists reasoning alongside general text generation among its intended strengths. That framing puts K2-Horizon in the same competitive lane as other large open MoE systems that have leaned on sparse activation to scale parameter counts while holding inference latency down.

Why it matters

Sparse MoE models have become the default strategy for teams that want frontier-scale capacity on a realistic serving budget. A 375B model that runs like a 23B one at inference time is far more approachable for organizations that can host it themselves, and open weights let them fine-tune and deploy without an API dependency.

A few practical notes for anyone evaluating it:

• Total vs. active parameters: 375B total, ~23B active per token.
• Architecture: mixture-of-experts, aimed at text and reasoning tasks.
• Licensing: listed as "other," so teams should review the terms on the model page before commercial use.

As with any fresh release, the real test will be independent benchmarking and community fine-tunes. For now, the model is live and downloadable, and the details on the Hugging Face repository are the authoritative reference.