Step 3

Step 3 is the flagship foundation model from StepFun, a Shanghai AI lab founded by former Microsoft Research Asia chief scientist Jiang Daxin and backed by Tencent. Released in July 2025, it is a sparse mixture-of-experts vision-language model with 321 billion total parameters (316B language backbone plus ~5B of vision components), of which only 38B activate per token. It supports native multimodal reasoning over images and text with a 65,536-token context window, and its weights were released openly.

The honest picture: Step 3 was quickly followed by newer siblings — Step 3.5 Flash (Feb 2026) and Step 3.7 Flash (May 2026, Apache 2.0, 196B total/11B active, 256K context) — which are faster, cheaper to run, and explicitly optimized for agentic coding workflows. Step 3 itself is the heavier, more general predecessor. Local deployment needs serious hardware (8–16 H20-class GPUs per the deploy guide), so most developers will use it through StepFun's cloud API or partners like OpenRouter and NVIDIA NIM.