Upstage launches Solar Mini 4 for efficient enterprise deployments

Upstage has introduced Solar Mini 4, a compact large language model that activates only 3 billion (3B) of its total 35 billion (35B) parameters during inference. The South Korean AI startup said its Mixture of Experts architecture reduces computational load while maintaining strong performance for enterprise use cases.

Solar Mini 4 recorded an overall score of 24 on Artificial Analysis Intelligence Index (AAII) v4.3.2, which Upstage said was the highest result among models at the 3B active parameter scale. The company said the model surpassed Alibaba’s Qwen 3.6-35B-A3B and Nvidia’s Nemotron 3.5 Lightning, and matched or outperformed higher-tier models including Nvidia’s Nemotron 3 Ultra, Google’s Gemma 4 31B, and Xiaomi’s MiMo V2.5.

With quantization, Solar Mini 4 can run on a single GPU, positioning it for on-premises deployments and large-scale repetitive task processing. Upstage is reorganizing its Solar LLM lineup into Solar Pro for complex agent work and Solar Mini for faster, token-efficient automation. The model is available through OpenRouter, with a 70% API pricing discount through the 10th.