Step-3.7-Flash

Step-3.7-Flash is a large language model released and open-sourced by StepFun (阶跃星辰), a portfolio company of Qiming Venture Partners, in early June 2026 . It represents the company's next-generation "Flash" model line, positioned specifically for the production deployment of AI agents rather than as a lightweight, cost-cutting variant .

The model is designed around a system-level optimization for agent workflows, coding, search, and multimodal tasks, aiming to balance speed, cost, reliable execution, and complex task completion . This reflects StepFun's broader strategic bet that the agent era introduces a third dimension to the intelligence-cost tradeoff: speed becomes critical because users no longer read model outputs in real time but instead wait for task completion, making latency directly felt .

Step-3.7-Flash follows the Step 3.5 Flash series, which in March 2026 briefly topped OpenRouter's global API call volume rankings, suggesting the product line has gained commercial traction . The 3.5 version featured a 196-billion-parameter sparse MoE architecture with ~11 billion activated parameters per token, MTP-3 multi-token prediction, and a hybrid attention mechanism handling up to 256K context windows . StepFun indicated that further technical details, demos, and ecosystem plans for 3.7-Flash would follow the initial release .