Step 3.5 Flash is an open-weights foundation model from the Shanghai-based lab StepFun, introduced in early 2026 as a counterweight to the industry's largest dense systems. Rather than scaling raw parameter counts, StepFun built the model around a sparse Mixture-of-Experts architecture that activates only the relevant specialists for each input, a design the company describes as emphasizing "intelligence density" and architectural efficiency. The same third-party coverage frames the release as a deliberate move toward models that can reason and execute tasks autonomously, positioning Step 3.5 Flash for workloads that go well beyond simple chat interactions.
Independent benchmarking on Artificial Analysis placed the original Step 3.5 Flash release below average on intelligence among compared models, while awarding top marks for speed and noting that the open-weight pricing remained reasonable relative to peers of similar size. The same benchmark page has since marked the 0202 build as deprecated and points users to a refreshed Step 3.5 Flash 2603 successor, which is the variant now recommended for evaluation. In practice, the model fits well for developers who want an open-source Chinese model that prioritizes fast inference, agent-style reasoning, and tool-driven execution, especially where cost efficiency matters more than peak leaderboard intelligence scores.
