StepFun’s Compact AI Model Outperforms Larger Rivals in Key Benchmarks
Shanghai-based StepFun has launched Step 3.5 Flash, a compact AI model that challenges China's heavyweight competitors by prioritizing efficiency over sheer scale. Despite having fewer parameters than rivals like Kimi K2.5 (1 trillion parameters) and DeepSeek V3.2 (1 trillion parameters), Step 3.5 Flash outperforms them in advanced reasoning, coding, and agent-based tasks.
The model achieved top results on four reasoning benchmarks, including AIME 2025 and IMOAnswerBench, surpassing offerings from DeepSeek, Moonshot AI, Zhipu AI, and MiniMax. Only Microsoft's OpenAI remains competitive. StepFun's breakthrough demonstrates that smaller, optimized models can deliver superior performance in critical AI applications.