Upstage said on Wednesday it has unveiled its in-house small large language model (LLM) Solar Mini 4.

Solar Mini 4 is a 35 billion-parameter model that applies a mixture-of-experts (MoE) structure. In actual inference, it selectively activates 3 billion parameters to reduce computational load.

In the Intelligence Index (AAII) by global AI model evaluator Artificial Analysis, it scored 24 points overall. Upstage said it recorded a high level of performance among models with 3 billion active parameters.

The company said it can run on a single graphics processing unit (GPU) when quantization is applied, making it usable in on-premise environments where a model is installed on internal corporate servers. It said it focused on lowering inference costs needed for repetitive data processing or work automation by using a small number of active parameters.

Upstage plans to operate the Solar LLM lineup by dividing it into Solar Pro and Solar Mini. It set Solar Pro for agent tasks that require complex judgment and Solar Mini for repetitive, high-throughput work automation as key use cases.

Solar Mini 4 is also provided as an API through the API routing platform OpenRouter. Upstage previously listed Solar Pro 4 on OpenRouter in August.

Upstage CEO Sung-hoon Kim (김성훈) said, "We systematized complex agent tasks as 'Pro' and large-scale work automation as 'Mini'." He added, "We will enable companies to choose an appropriate AI model depending on their work characteristics and throughput."