From the source
Only 3B active parameters per token, with leading performance in its class and lower costs for repetitive agent workload We launched Solar Mini 4, an agent-optimized model pretrained from the ground up by Upstage.
It activates just 3B of its 35B total parameters per token while achieving the highest overall score among the 3B-active models in the Artificial Analysis Intelligence Index comparison.
Solar Mini 4 is designed to handle repetitive tasks—including information retrieval, structured output generation, and tool use—using less compute and at lower cost.
Key capabilities Highest overall performance among the 3B-active models evaluated : Solar Mini 4 scored 24.1 on the Artificial Analysis Intelligence Index v4.3.2, the highest result among the models with 3B active parameters included in the comparison.
Optimized for agent workloads : Solar Mini 4 scored 22.3% on AutomationBench-AA, which evaluates agents on real SaaS workflows, and 47.2 in a separate τ³-Banking evaluation.
Efficient MoE architecture : The model has 35B total parameters while activating only 3B per token.
512K context window : Supports up to 512K tokens of context and up to 128K output tokens.
…
