Upstage's Solar Open2 arrives as a 250B MoE model

Upstage has released Solar Open2 250B, a text-generation model built on a mixture-of-experts (MoE) architecture and made available on Hugging Face. The release marks a notable step up in scale for the Korean company's open Solar line, which has focused on delivering competitive language models outside the largest US labs.

The headline figure is a 250-billion-parameter footprint. As an MoE system, the model routes each token through a subset of specialized expert networks rather than activating every parameter at once — an approach that lets developers pursue large total capacity while keeping inference costs more manageable than a comparably sized dense model.

Bilingual by design

Solar Open2 is aimed squarely at English and Korean workloads. That bilingual focus has been a consistent thread for Upstage, whose earlier Solar models earned attention for punching above their size, and it positions the release as a practical option for teams building products for Korean-language users where many Western open models underperform.

A few details worth noting from the release:

• Distributed openly under a custom ("other") license, so teams should review the terms before commercial use.
• Weights are hosted directly on Hugging Face for download and self-hosting.
• Context length and detailed benchmark figures were not specified in the initial listing.

Why it matters: Large open MoE models remain relatively rare, and most of the highest-profile ones come from a handful of labs. A 250B bilingual release from Upstage widens the field and gives Korean-focused developers a large-scale option they can inspect and run themselves rather than access only through an API.