‘Punches above its weight’: compact AI model from China’s StepFun outshines larger rivals

Chinese artificial intelligence start-up StepFun has unveiled a lightweight AI model that it says punches above its weight, rivalling larger systems from domestic competitors including DeepSeek and Moonshot AI as competition intensifies in the country’s AI sector.

The Shanghai-based AI lab said on Monday its latest Step 3.5 Flash model was designed to deliver advanced reasoning and agentic capabilities while maintaining efficiency.

Despite its relatively modest size of about 196 billion parameters – far smaller than Moonshot AI’s Kimi K2.5 with 1 trillion parameters or DeepSeek V3.2 with 671 billion parameters – Step 3.5 Flash outperformed its larger rivals across several benchmark tests measuring agentic, reasoning and coding capabilities, according to the company’s self-reported results.

Parameters are the variables that encode an AI system’s “intelligence”, with a larger number usually indicating stronger performance.

Step 3.5 Flash topped four reasoning benchmarks, including AIME 2025 and IMOAnswerBench, outperforming leading systems from DeepSeek, Moonshot AI, Zhipu AI and MiniMax, and trailing only Microsoft-backed OpenAI in certain tests.

StepFun says its lightweight model was designed to deliver advanced reasoning and agentic capabilities while maintaining efficiency. Photo: Handout

The model’s compact size and focus on reasoning were deliberate choices, according to Zhu Yibo, StepFun’s co-founder and chief technology officer. Zhu said the team had prioritised “strong logic capability, efficient context window and fast speed” when developing the new system, which was purpose-built for the AI agent era.

South China Morning Post