StepFun (阶跃星辰)

Overview

StepFun (阶跃星辰) is a Shanghai-based AI lab whose Step series of models has climbed domestic leaderboards and expanded into multimodal and agent workloads. The flagship Step 3.7 Flash uses a sparse MoE architecture (roughly 198B total / 11B active parameters) with native image and video understanding, tool calling, and a 256K context window tuned for agentic coding and planning. A free quota is available to new platform users, and the API is OpenAI-compatible, so it drops into existing agent stacks.

Beyond text, StepFun ships Step1X-Edit for image editing (style transfer, inpainting, text replacement) and StepAudio for speech synthesis and real-time voice. The lab has also pushed toward on-device and terminal AI with its Step AOS agentic OS. In our evaluation, Step 3.7 Flash stands out for high-throughput, multimodal agent loops where a single call can plan, reason, and emit code or structured output — a good fit for builders who want open-ish, cost-effective intelligence with vision built in.

Key Features

• Native Multimodal — Step 3.7 Flash understands images and video directly, no separate vision model needed in agent loops.
• Agent-Optimized — MoE architecture delivers fast, low-latency inference for multi-step tool-use and planning.
• OpenAI-Compatible API — Works with standard SDKs and popular coding agents out of the box.
• Open Ecosystem — Open-weight models plus Step1X-Edit and training frameworks on GitHub.
• Long Context — 256K-token window supports long documents, codebases, and conversation history.

Pricing

Comparison

vs. DeepSeek: Both are Chinese open-weight labs with cheap APIs; DeepSeek has a larger global community, while StepFun leads on native multimodal and agent-tuned models. vs. Kimi: Kimi emphasizes long-context chat and document work; StepFun adds built-in vision and tool-use for agent pipelines. vs. Qwen: Qwen offers the broadest open-model range, but StepFun’s Flash models are tuned specifically for real-time agent performance.