Step 3.7 Flash: Achieving Flagship-Level Performance at One-Ninth the Cost of Claude Opus 4

In 1492, when Columbus set sail across the Atlantic, the success of his voyage wasn’t driven by sheer romanticism alone. Freshwater, food supplies, ship integrity, and rigging durability all played crucial roles. Later, the Dutch-designed “Fluyt” merchant ships revolutionized oceanic trade with their low construction costs, spacious cargo holds, and stable round-trip capabilities. This engineering ingenuity transformed oceanic trade from solitary explorations into a repeatable and scalable business venture.

The current competition among AI models stands at a similar crossroads. In the past, discussions about models revolved around parameter counts, benchmark scores, and peak capabilities. However, as AI agents transition into fully operational production environments, the true challenges are shifting. The key questions now are: Can the model handle high-frequency requests consistently? Can it reliably call tools? Can it be seamlessly integrated into existing enterprise workflows for long-term use? These questions often aren’t addressed by traditional benchmark lists.

Amid this turning point, Chinese AI company StepFun has introduced Step 3.7 Flash. With task costs reduced to approximately one-ninth of Claude Opus 4’s, this production-grade Flash model is optimized for agents, coding, search, and multimodal workflows. Recently released and open-sourced, it represents a new frontier in AI model design.

The Changing Role of Flash Models

Traditionally, Flash models were perceived as “lightweight versions of flagship models.” Their sole selling points were speed and affordability. However, as agents become central to workflows, the requirements for Flash models have transformed drastically.

If a model frequently deviates from its objectives across multiple tasks, neither companies nor individuals can adopt it with confidence. On the other hand, if it achieves a balance of speed, cost, tool integration, multimodal understanding, and ecosystem compatibility, it can serve as a reliable foundation for agent systems.

In essence, the Flash model demanded by the agent era is evolving from “a faster, smaller model” to “the most efficient production-grade base model.” This new breed of Flash models not only approaches the performance limits of flagship models but also withstands the pressures of large-scale agent calls. Step 3.7 Flash is positioned precisely as the latter—a next-generation agent-oriented base model.

Native Multimodal Understanding Unlocks

Practical Adaptability

For agents operating in production environments, the first hurdle is accurately comprehending the actual work setting. Real-world agent tasks are distributed across complex UIs, documents, chart systems, browser pages, specialized software, and internal tools. Models focused solely on text-based Q&A struggle to address these challenges effectively.

Step 3.7 Flash emphasizes native multimodal understanding and execution capabilities. It can comprehend UIs, charts, documents, images, and application interfaces. Furthermore, for complex visual issues, the model autonomously crops, zooms, and rereads images. When encountering uncertain information, it initiates its own searches, cross-verifying text and image data.

This represents a counterintuitive design philosophy. For a Flash model with 11B active parameters, cramming vast visual knowledge into its weights isn’t ideal. Instead, Step 3.7 Flash adopts an opposite approach: retaining only the core inference engine within its weights while extrapolating perceptual boundaries and world knowledge during inference. By rapidly “looking multiple times and investigating multiple times,” the model compensates for areas where its parameters might fall short. Low latency and high throughput aren’t merely deployment advantages—they’re intricately built into the model’s capabilities.

Practical Use Cases

In a cockpit operation demo, users simply input “how to take off,” and the model autonomously selects the cockpit region, identifies instruments and buttons, understands the operational logic of the current interface, and generates a step-by-step tutorial. This ability to transform a crowded, unfamiliar visual environment into actionable task guides surpasses mere image recognition.

Another intriguing demo showcases integration with smartphone GUI agent flows. By connecting a smartphone via USB and sending screenshots to the model, it interprets ongoing screen activity. For instance, when shown the trending rankings on a reading app, the model not only reads the text but also grasps the ranking structure. It identifies which items represent book titles, covers, or current ranking positions, enabling the next operational steps.

The model also proves useful in business processes requiring complex understanding. On pages displaying user evaluations, image evidence, business responses, and action buttons simultaneously, the model organizes who filed a complaint, what the dispute is about, and the evidence presented before making an appropriate decision. Interfaces mixing text, images, judgments, and operational entries are precisely the environments multimodal agents are meant to address.

Enhanced Search Capabilities Bolster Reliability

Another key enhancement of Step 3.7 Flash lies in its network connectivity and visual search capabilities. Real-world tasks often involve dynamic information, external resources, multi-source evidence, and incomplete inputs. If a model relies solely on its internal knowledge, it risks failures in timeliness and accuracy.

For instance, when shown a building image, the model first reads visible clues from the image, generates search terms centered on those clues, performs external searches, and combines the visual and textual information to craft a complete answer. The searches go beyond merely returning a list of web links—they actively explore, filter, cross-check, and organize evidence around the task goal. This methodology aligns with the actual needs of Search and Research agents.

According to official reports, Step 3.7 Flash demonstrates flagship-like performance on complex visual task benchmarks such as SimpleVQA Search and V* (Python). Its ability to proceed with tasks amidst incomplete information and reduce unverified answers highlights its practical utility.

Sparse MoE Architecture Enables Exceptional

Scalability

The key difference between agents and regular chatbots lies in their high call density. While typical question-and-answer interactions often end after one exchange, agents need to repeatedly observe environments, call tools, and interpret results to complete tasks. Coding agents read code, modify files, and execute commands. Search agents look for, cross-check, and organize information. As call frequency increases, a model’s speed and cost can become bottlenecks for the entire system.

Step 3.7 Flash employs a Sparse MoE (Mixture of Experts) architecture. Although the total parameters amount to 196B plus 1.8B ViT, active parameters are kept to just 11B. This design enables a maximum generation speed of 400 tokens per second and allows simultaneous operation of 40 agents.

This scalability means high-frequency agents, coding agents, search agents, multimodal agents, and enterprise knowledge work agents can complete more observations, calls, and inferences within the same timeframe. The model can run 40 virtual personas in parallel, enabling them to make multifaceted product evaluations, among other large-scale agent cluster tasks.

”Fast and Cheap” No Longer Suffices

The emergence of Step 3.7 Flash highlights a subtle shift in AI model competition. The focus is transitioning from the “age of adventurers,” where peak performance was the ultimate goal, to the “age of merchant ships,” emphasizing cost, stability, and scalability.

As agents become deeply integrated into enterprise workflows, the criteria for models are shifting from achieving top scores to delivering consistent results across millions of daily calls, understanding multimodal real-world environments, and seamlessly interacting with existing tools and interfaces. These foundational capabilities are likely to define AI model competitiveness in the next phase.

By reducing task costs to one-ninth that of Claude Opus 4 while approaching flagship-level performance, Step 3.7 Flash embodies the design philosophy of a “merchant ship.” Its open-source release further positions it as a pivotal development in the infrastructure models of the agent era.