Models & Weights

Solar Open 2

Apache-2.0 in every clause but two: name your model 'Solar', and say so in public

Open code / restricted weightsModel

What it is

Upstage's sovereign-AI flagship: a 250B-total / 15B-active hybrid-attention MoE trained on ~12T tokens on NVIDIA B200s, with 1M-token context and an agentic-use focus. Korea's biggest open-weights swing to date, positioned to run on as little as two GPUs quantized, with NVFP4 and community GGUF builds circulating.

Why it's interesting

It's been on our watch list since July as a 'custom Solar License' curiosity, and the traction — 748 HF likes, ~29k downloads a month — says it's time to read the license. The read is instructive: an Apache-2.0 chassis granting perpetual, royalty-free, commercial-inclusive rights with no MAU caps and no AUP, then two bespoke conditions bolted on — derivative models must be named with a 'Solar' prefix, and public-facing materials must prominently display 'Built with Solar'.

Use cases

  • Self-hosted agentic workloads at frontier-adjacent scale on modest GPU counts
  • Korean-language and sovereign-deployment scenarios where provenance matters
  • A case study in branding-locked licensing on an otherwise permissive base

Who it's for

Enterprise self-hosters, sovereign-AI watchers — and anyone naming a fine-tune

Setup

Advanced. Multi-GPU node for full precision; Upstage claims two-GPU serving quantized, with NVFP4 and community GGUF builds available

Limitations & cautions

The branding conditions propagate to every derivative regardless of scale, the bespoke license means legal review no standard text would need, and the two-GPU serving claim is Upstage's own quantized best case — plan for more.

Editorial takeaway

The friendliest fine print we've read all summer — right up until it names your model for you. Two clauses are all it takes to leave the OSD behind.

Related & alternatives