StepFun Trillion-Parameter MoE Enterprise Agent

📌 Background

In 2026, the large model industry shifted from a technical race to commercial realization, leading to an explosion in demand for enterprise-grade documents, code, and Agents. StepFun leverages a trillion-parameter MoE architecture to reduce inference costs. In early 2026, it completed a B+ round of financing exceeding 5 billion RMB, setting a new single-deal record in the domestic large model sector, with cumulative financing reaching nearly 2.5 billion USD (company-disclosed figures, unverified). The company holds a leading market share in private deployments for finance, energy, and manufacturing, and is currently advancing its Hong Kong IPO.

👤 Target Customers

Large and medium-sized enterprises in finance, energy, manufacturing, and government sectors, as well as developers requiring API access and overseas voice model clients.

💰 Revenue Streams

1) Enterprise private deployment and customized contracts (charged by seat, usage, or custom service); 2) Large model API usage revenue (Step series models billed by token or prompt); 3) Licensing for terminal device models and revenue sharing from the core module ecosystem.

🧮 Cost Structure

Core costs include training compute for the trillion-parameter MoE model and maintenance of inference clusters. The company collaborates with domestic chip providers like Huawei Ascend and Enflame to reduce per-inference compute consumption, alongside costs for overseas commercial teams and delivery personnel.

🛡️ Moat

Engineering capabilities in cost reduction via sparse activation of trillion-parameter MoE models, with Step 3.7 Flash achieving inference speeds of 409 Tokens/s with only ~11B active parameters. This is bolstered by domestic chip ecosystem barriers and a first-mover advantage in private deployments for finance and manufacturing.

🔑 Keys to Success

• Reducing trillion-parameter model inference costs to an enterprise-acceptable range via MoE sparse activation
• Penetrating high-paying industries through benchmark private deployment cases in finance and manufacturing
• Building a compute cost barrier and promoting terminal distribution in partnership with domestic chip manufacturers

⚠️ Risks

• Continuous decline in API unit prices driven by open-source models makes maintaining high premiums for closed-source models difficult
• Risk of commercial viability being questioned post-IPO if phenomenal applications fail to emerge
• Overseas model API sales subject to geopolitical and compliance uncertainties

🏢 Cases

• StepFun Step-2 Trillion-Parameter MoE Model
• StepFun Step 3.7 Flash Open-Source Lightweight Model
• StepFun AI Xiao Cai Shen Financial Application

📊 SWOT Analysis