Solar Mini 4

Solar Mini 4 is a compact language model developed by Upstage AI, optimized specifically for high-volume agentic workflows, multi-turn tool calling, and structured task automation. Designed as a cost-efficient alternative within Upstage's model ecosystem, it balances high computational efficiency with strong performance across long-context reasoning tasks.

Architecture and Size

The model utilizes a mixture-of-experts (MoE) architecture with a total of 35 billion parameters. To maximize inference speed and reduce operational overhead, Solar Mini 4 selectively activates only 3 billion active parameters per token during runtime. It supports an extensive 524K context window alongside a large output token capacity, allowing it to process lengthy reports and maintain extended conversational sessions without segmentation.

Capabilities

Solar Mini 4 is engineered for agent-based application use cases, data parsing, and multi-step reasoning. It features native support for fluent Korean alongside robust proficiency in English and Japanese. Its design focuses on lowering per-token expenses while maintaining high accuracy for repetitive administrative tasks, function calling, and systematic policy checking.