Upstage says Solar Mini 4, designed for repetitive agent workloads, activates 3 billion of its 35 billion parameters per token and supports a 512,000-token context, structured output and parallel tool calls. The minimum deployment for the quantized model is one 80GB H100, but the company’s measurement of more than 70 tokens per second per request at 32 concurrent requests used two GPUs, 4K-token input and 1K-token output. API list prices are $0.10 per million input tokens and $0.40 per million output tokens. Full-weight memory requirements and actual call counts also matter; maximum-context performance on a single GPU has not been demonstrated.