Step 3.7 Flash

Step 3.7 Flash is an open-weight model engineered for real-world applications, offering a unique blend of capabilities for dynamic environments. Key features include: • Native multimodal understanding and action • Enhanced web and visual search • Reliable tool use and orchestration • Broad agent ecosystem compatibility • High-efficiency processing at up to 400 tokens per second (TPS) This model excels at interpreting diverse visual information, from product interfaces and documents to charts and natural scenes. It can then translate these perceptions into actionable code or precise tool commands, allowing it to interact with and operate various systems effectively. Its advanced search functions extend beyond traditional web queries, delving deeper into sources and recognizing emergent concepts in visual data that other systems might overlook. Step 3.7 Flash demonstrates robust control over external tools such as terminals, browsers, and office applications, maintaining coherence and reliability even during extended operations. This minimizes common issues like toolcall breakdowns or failed runs. With a context window of 256K and approximately 11 billion active parameters, it provides a powerful foundation for complex tasks. Designed for seamless integration, Step 3.7 Flash is compatible with mainstream agent harnesses and skill frameworks like Claude Code, KiloCode, Hermes Agent, and OpenClaw. This ensures lower integration costs and less disruption to existing development workflows, making it ideal for engineering teams and developers building sophisticated, interactive systems that require perception, planning, and execution.