Upstage: Solar Mini 4

Solar Mini 4 is Upstage's compact, cost-efficient language model in the Solar LLM family, positioned alongside Solar Pro for agentic and automation workloads. It uses a mixture-of-experts design with 35 billion total parameters and about 3 billion active parameters per token, which Upstage targets for fast inference and lower serving cost while keeping strong multilingual chat quality in Korean, English, and Japanese.

The model supports up to 512K tokens of context, up to 128K output tokens, optional reasoning modes, structured JSON outputs, and tool calling. It is offered through the Upstage Console and Playground APIs, OpenRouter, and on-premises deployment; weights are proprietary and not published on Hugging Face. Training data runs through a February 2026 cutoff.

On third-party evaluations run by Artificial Analysis (Intelligence Index v4.3.2, reasoning configuration), Solar Mini 4 scores 24 on the composite index and reports about 25.8% on Humanity's Last Exam, 1090 on GDPval-AA, and 22.3% on AutomationBench-AA. Upstage markets it for high-volume agent steps, multilingual customer workflows, and pairing with Solar Pro for harder reasoning tasks.