Step 3.7 Flash

Step 3.7 Flash is a new artificial intelligence model designed to make digital agents smarter and more efficient. It is built to understand both text and images while performing real-world tasks like searching the web, using software tools, and writing code. This model is part of the Step 3 series and represents a major upgrade in how computers can process visual information and take action based on what they see.

Benefits

Step 3.7 Flash offers several key advantages over older models. It can understand complex images such as product interfaces, documents, charts, and natural scenes. When it sees something, it can write code or call tools to act on that information. This makes it very strong at visual reasoning tasks where it performs as well as models that are five times its size. The model also improves web search by finding more sources and performing deeper follow-ups. It recognizes specific details and new concepts that other systems might miss. Additionally, it uses tools like terminals, browsers, and office software with high reliability, reducing the chance of failed tasks. It is compatible with many existing agent platforms, which lowers the cost of integration. For coding tasks, it shows significant improvements in solving software problems and running terminal commands. It also supports an Advisor Mode that uses a larger model only when needed, making it much cheaper to run while still delivering high performance.

Use Cases

This tool is ideal for businesses and developers who need automated agents to handle complex workflows. It can be used in finance, accounting, and data analysis to perform tasks that require deep domain knowledge. Companies in manufacturing and engineering can use it for production scheduling and analyzing technical data like heat treatment processes. Developers can leverage its coding capabilities to build software faster and fix bugs more effectively. It is also useful for research tasks where deep retrieval and information synthesis are required. The model works well with visual tools to help with tasks like identifying objects in images or understanding complex diagrams. It can be deployed in various environments including cloud servers, data centers, and local machines with high memory.

Pricing

The specific pricing for Step 3.7 Flash is not publicly listed in the available information. However, the model is available through the StepFun Open Platform API. It supports flexible deployment options that may allow users to choose a pricing model that fits their needs, such as pay-per-use or subscription plans. The Advisor Mode mentioned in the documentation highlights a cost efficiency of roughly one-ninth compared to other high-performance coding models, suggesting that the standard pricing is designed to be competitive for enterprise use.

Vibes

Public reception and detailed reviews are not available in the provided text. The announcement focuses on technical benchmarks and performance metrics rather than user testimonials. The model has achieved impressive scores in various testing environments, such as a 92.8% F1 score on DeepSearchQA and 56.3% on SWE-Bench Pro. These numbers indicate strong performance among technical experts and developers who test AI models. The emphasis on efficiency and cost-effectiveness suggests that the community may view it as a practical solution for scaling agent applications without breaking the bank.

Additional Information

Step 3.7 Flash is developed by StepFun and is available via their Open Platform API, web interface, and mobile app. It supports deployment across cloud, data center, and local environments. The model requires specific hardware to run efficiently, including systems with at least 128GB of unified memory such as NVIDIA DGX Station, AMD Ryzen AI Max+ 395-based systems, or Mac Studio and Book Pro. It is compatible with various inference frameworks like vLLM, SGLang, and Hugging Face Transformers. The model has 196 billion total parameters but only uses 11 billion active parameters during inference, which contributes to its efficiency. It is designed to work seamlessly with mainstream agent harnesses like Claude Code, KiloCode, and Hermes Agent.