stepfun-ai/step-3.7-flash

Step 3.7 Flash is an AI model released by StepFun, a Chinese AI company. It is designed to process both text and images at the same time, and it targets developers building automated workflows that need to handle large amounts of work quickly. The model is big by design: it has 198 billion total parameters but activates only about 11 billion of them per token, which lets it run faster than its full size would suggest. The model supports a context window of 256,000 tokens, meaning it can read very long documents in a single pass. It also offers three reasoning depth settings, labeled low, medium, and high, so a developer can trade off speed against thoroughness depending on the task. At maximum throughput the model can produce up to 400 tokens per second. StepFun positions this model for agentic tasks: situations where an AI needs to call external tools, browse multiple sources, verify its own findings, and complete multi-step jobs without human supervision at each step. The README describes use cases like parsing long financial reports, running search-and-verify loops across many sources, and operating multiple coding agents in parallel. It also scores well on benchmarks for visual understanding of user interfaces and data charts, and for following tool-use instructions without drifting from the given constraints. For code tasks, the model can read multi-file repositories, find bugs from issue descriptions, and write patches that pass automated tests. Its benchmark results place it solidly in the upper tier for software engineering tasks, though the README acknowledges specific areas where it falls short of the absolute top scores. Pricing runs at $0.20 per million input tokens on a cache miss, $0.04 per million on a cache hit, and $1.15 per million output tokens. The model is available through StepFun's own API platforms (one for global users, one for China), as well as through OpenRouter and NVIDIA NIM. For local use, it requires a machine with at least 128 GB of unified or GPU memory. It works with common open-source serving tools including vLLM, SGLang, and llama.cpp.