Step 3.7 FlashSee, code, and act at lightning speed with open AI agents

Step 3.7 Flash

Step 3.7 Flash is a high-efficiency multimodal AI model designed for real-world agentic applications. Built on a 196B parameter architecture with 11B active parameters, it delivers exceptional performance in coding, reasoning, and visual understanding while maintaining Flash-level speed and cost efficiency.

Product Highlights

• Agentic Coding Excellence: Achieves 56.3% on SWE-Bench Pro and 59.6% on Terminal-Bench 2.1, outperforming comparable Flash models from DeepSeek and Gemini while approaching larger Pro-tier systems.
• Native Multimodal Understanding: Processes and acts upon images, documents, charts, and natural scenes with integrated tool use for comprehensive visual reasoning.
• Advanced Search Capabilities: Scores 75.8% on BrowseComp and 47.2% on HLE with tools, enabling deep research and multi-source information synthesis.
• Reliable Tool Orchestration: Drives terminals, browsers, Office tools, and search systems with minimal drift and failed tool calls across long-horizon workflows.
• Agent Ecosystem Compatibility: Works seamlessly with Claude Code, KiloCode, Hermes Agent, OpenClaw, and OpenCode with reduced integration costs.

Use Cases

• Software Development: Autonomous coding agents that write, debug, and deploy code across heterogeneous development environments.
• Enterprise Automation: Long-horizon task execution combining document processing, data analysis, and cross-application orchestration.
• Research and Analysis: Deep search and synthesis across academic papers, technical documentation, and live web sources.
• Visual-GUI Interaction: Phone and desktop automation through GUI perception, clicking, and verification across multiple applications.

Target Audience

Step 3.7 Flash serves developers building AI-powered applications, enterprise teams automating complex workflows, and researchers requiring efficient multimodal reasoning—delivering near-Pro performance at Flash-tier economics.