Step 3.7 Flash – Open-source multimodal model for speed and agents (static.stepfun.com)

Step 3.7 Flash has been announced as a cutting-edge open-source multimodal model designed to enhance efficiency for real-world AI agents. This model features native multimodal understanding, enabling it to interpret a wide array of visual inputs—from product UIs to complex documents—and then either write code or invoke tools based on its observations. Key enhancements include a significant boost in multimodal search capabilities, allowing deeper and broader searches compared to previous models, and advancements in long-horizon task execution, enabling the model to autonomously manage assignments across various digital interfaces. Significantly, Step 3.7 Flash integrates seamlessly across popular agent ecosystems, reducing integration costs and increasing workflow efficiency. With its Advisor Mode, the model intelligently collaborates with a larger advisor model during complex tasks, achieving impressive coding performance at a fraction of the cost compared to leading alternatives. Performance benchmarks further illustrate its prowess, surpassing competitors in multimodal tasks and search-heavy evaluations. By focusing on bridging comprehension and action, Step 3.7 Flash not only represents a technological leap in AI multimodal capabilities but also establishes a new standard for efficiency and versatility in the AI/ML community, paving the way for more robust, intelligent agents in professional environments.