Chinese artificial intelligence start-up StepFun has unveiled a compact AI model, Step 3.5 Flash, that it says can rival far larger systems from domestic competitors, sharpening competition in China’s fast-moving AI sector.
The Shanghai-based company positions the new release as proof that efficiency, not sheer scale, can drive performance in advanced reasoning and agent-based tasks.
StepFun’s model defies size with reasoning performance
The StepFun company’s Step 3.5 Flash AI has many fewer parameters than some of their competitors’ products that produce Kimi K2.5 (1 trillion parameters) and DeepSeek V3.2 (1 trillion parameters).
However, Step 3.5 Flash has proven to be more useful than those larger AI models in several of the benchmarks performed, particularly in the areas of reason, computer coding, and the ability to create intelligent agents (agents that can perform tasks with varying levels of intelligence).
Specifically, the Step 3.5 Flash AI model achieved the best results on four different reasoning benchmarks (AIME 2025; IMOAnswerBench; etc.), beating all other models in both the DeepSeek, Moonshot AI, Zhipu AI, and MiniMax product lines – with the only competitors being Microsoft’s OpenAI.
Kimi K2.52, DeepSeek V3.2, and many of the other AI models in the marketplace are struggling to compete with the performance of Step 3.5 Flash. Parameters are the individual pieces of information used in a model that form the basis of a model’s “intelligence,” and large numbers of parameters in an AI model are generally associated with improved performance on tasks.
Prioritizing logic, speed, and agents
The model’s design is intentional and consists of built-in trade-offs, according to StepFun. CTT, CTO and co-founder Zhu Yibo said their focus on practical uses was more important to them than mere headline size, stating “logic capability, context window size, and speed were what mattered most to us,” in describing the overall purpose of the system for use in an AI-agent dominated world.
According to Zhu, they designed the model to be based off of their experiences with earlier, larger models, as larger models took more time to train and, as a result, usually took longer than expected before being delivered.
Zhu also said on an earlier post that “the strength of Step 3.5 Flash comes from the same area we care about most: its agent-based behaviour that will enable logic reasoning to happen effectively.”
A number of Chinese semiconductor companies, including Huawei and MetaX, have modified the hardware for their chips to be compatible with StepFun’s new hardware architecture, demonstrating confidence in the efficient operation of this system.
The timing of this launch is coincident with an increase in the number of new products and prototypes that will be demonstrated during the Lunar New Year and as Chinese companies develop or release new technology.
Alibaba and Moonshot AI have also announced that they will release updates on their models; however, Zhipu and MiniMax are still working on developing their products.
Recently, StepFun secured 5 billion RMB (approximately 720 million USD) in funding during a B++ funding cycle. Included in this round of investment were government-funded agencies and other well-known firms; this illustrates the increasing confidence that small, fast models have the potential to dominate the Chinese market for AI.
If you’re reading this, you’re already ahead. Stay there with our newsletter.
Source: https://www.cryptopolitan.com/stepfuns-ai-tests-the-bigger-is-better-guess/
