Step-2
by Step Fun·China·Released
StepFun's trillion-parameter MoE flagship — one of China's most-watched closed labs.
About this model
Step-2 is StepFun's flagship — a trillion-parameter MoE that placed the lab firmly in the Chinese frontier conversation alongside DeepSeek and Moonshot. StepFun (founded by ex-Microsoft Research Asia executive Lin Junshan in 2023) has raised over $1B and operates a particularly large international serving footprint via partnerships.
Closed weights, accessed via StepFun's platform. The lab has shipped specialist variants for math (Step-Math), vision (Step-Vision), and audio (Step-Audio). Among the more research-active Chinese labs — Step-Reasoner is one of the more credible open reasoning-trace contributions outside DeepSeek R1.
Strengths
- •Trillion-parameter scale with frontier-competitive quality
- •Strong specialist variants (math, vision, audio)
- •Backed by ex-MSRA leadership
- •200K context window
Limitations
- •Closed weights — no fine-tuning or self-hosting
- •Less developer ecosystem than Qwen/DeepSeek/Kimi
- •International procurement friction (Chinese-origin closed model)
When to use it
- →Chinese-market enterprise deployments
- →Specialised math / multimodal workflows
- →Research collaboration with StepFun's published variants
Architecture & training
Architecture details limited; StepFun has confirmed trillion-parameter MoE without specifying active count. Training infrastructure is reportedly a mix of H100 and Huawei Ascend chips.