Step 3.7 Flash - Step Star's open-source next-generation Flash model

Step 3.7 What is Flash?

Step 3.7 Flash is a new generation of open-source Flash model launched by StepStar for the production stage of Agents. It adopts a sparse MoE architecture and has a maximum generation speed of 400 Tokens/s. The model is systematically optimized around Agent, Coding, Search and multimodal workflow, achieving a balance between speed, cost, reliable execution and the ability to complete complex tasks, providing developers with a high-efficiency foundation model for building production-grade Agents.

Step 3.7 Main Functions of Flash

• Native Multimodal Understanding and ExecutionIt natively understands UI, charts, documents, images, and application interfaces, transforming complex visual information into structured results, code generation, and executable tasks, and supports autonomous image cropping, zooming, and rereading.
• Enhanced Internet and Visual SearchEnhance network retrieval and image search, and proactively acquire and cross-compare multi-source evidence across text and images in an open information environment.
• High-reliability tool calling and orchestrationIn long-term, multi-round Agent workflows, ensure stable calls to APIs, browsers, terminals, Office tools, and external systems, maintaining consistent task trajectories.
• Agent ecosystem compatibility optimizationIt provides compatibility and adaptation for mainstream agent frameworks such as Claude Code, KiloCode, RooCode, OpenCode, Hermes Agent, and OpenClaw, as well as the MCP/Skills protocol.
• Dual deployment on local and cloudSupports cloud API calls, and is optimized for local deployment on personal workstations, providing a GGUF multi-precision client version.

Step 3.7 Technical Principles of Flash

• Sparse MoE architectureWith a total of 196B + 1.8B (ViT) parameters and only 11B activation parameters, it achieves a better balance between model capability and inference cost.
• Visual perception-search-reasoning fusionIt deeply integrates visual perception, search, and reasoning processes, and proactively initiates searches for cross-validation when faced with complex visual problems.
• Long-range mission trajectory maintenanceIn multi-round Agent workflows, optimization maintains consistency in task trajectories, reducing deviations and execution failure rates.
• Multi-precision end-side optimizationProvides GGUF format support to adapt to local deployment needs with different hardware resources.

How to use Step 3.7 Flash

• Cloud access
  • Domestic Open Platform APIVisit https://platform.stepfun.com to register an account and obtain an API Key, after which you can call model capabilities through the standard interface.
  • Overseas Open Platform APIAccess https://platform.stepfun.ai to connect to overseas nodes, suitable for international business scenarios.
  • Studio Online ExperienceVisit https://studio.stepfun.com/ to experience model dialogue and multimodal capabilities directly on the web without any configuration.
• Open source deployment
  • GitHub repositoryGo to https://github.com/stepfun-ai/Step-3.7-Flash to download the complete code and weights, and follow the documentation to deploy it locally or on a server.
  • Huggingface modelObtain the model file from https://huggingface.co/stepfun-ai/Step-3.7-Flash. It supports direct loading and inference for mainstream frameworks such as Transformers.
  • Modelscope ModelDownloadable from https://modelscope.cn/models/stepfun-ai/Step-3.7-Flash, suitable for domestic developers to quickly retrieve.
  • End-side GGUF versionVisit https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF to obtain the multi-precision quantization version, which can run smoothly on personal workstations and even low-configuration devices.
• Application and Ecosystem
  • Step AI AppDownload the Step AI mobile app to experience Step 3.7 Flash's question-answering and agent capabilities directly on your phone.
  • Agent framework integrationConfigure the Step 3.7 Flash API endpoint in verified Agent tools such as Kilo Code and Nous Research (Hermes Agent) to drive coding and automated workflows from the underlying model.

Step 3.7 Flash's Core Advantages

• Speed ReasoningA single request can reach up to 400 TPS, making it suitable for high-frequency, multi-round, and low-wait agent applications.
• Extremely low activation parametersWith only 11B activation parameters, inference costs are significantly reduced, making it extremely cost-effective.
• Production-grade reliabilityIt performs stably in real-world benchmarks such as Toolathlon (49.5%), ClawEval-1.1 (67.1%), and GDPval (45.8%).
• Multimodal deep understandingIt rivals larger-scale flagship models on complex vision tasks such as SimpleVQA (79.2%) and V* (95.3%).
• Eco-Plug and PlayPre-optimized for mainstream Agent frameworks and MCP/Skills protocols to reduce access costs.

Step 3.7 Flash project address

• Project official website: https://static.stepfun.com/blog/step-3.7-flash/
• GitHub repositoryhttps://github.com/stepfun-ai/Step-3.7-Flash
• HuggingFace model libraryhttps://huggingface.co/stepfun-ai/Step-3.7-Flash

Step 3.7 Comparison of Flash with Competitors

Step 3.7 Application Scenarios of Flash

• Coding AgentIt performs exceptionally well in SWE-Bench Pro and Terminal-Bench, and can stably perform development tasks such as code generation, debugging, and terminal operations.
• Multimodal knowledge workIt automatically understands UI interfaces, charts, and documents to generate structured analysis or operation guidelines, suitable for product analysis, design review, and software tutorial generation.
• Enterprise Process AutomationIt can be embedded into enterprise workflows to perform data entry, report generation, and cross-system operations by making stable calls to APIs, Office tools, and external systems.
• Visual search and verificationIt proactively initiates image searches and cross-validation when information is uncertain, making it suitable for e-commerce price comparison, content review, and fact-checking.