StepFun: China's AI Foundation Model Startup

StepFun: China's AI Foundation Model Startup | Generated by AI

Home 2026.07

Question: Introduction to StepFun (https://www.stepfun.com/)

Answer:

Overview

StepFun (Chinese name: 上海阶跃星辰智能科技有限公司) is one of China’s leading foundation model startups, often grouped with companies such as Moonshot AI, MiniMax, and Zhipu AI as the new generation of Chinese frontier AI labs. It was founded in April 2023 and is headquartered in Shanghai. (Wikipedia)

The company is building:

Large Language Models (LLMs)
Multimodal models (vision + language)
Speech models (ASR/TTS)
Image generation/editing models
Agent-oriented models
API platform for developers

Their positioning today is similar to what OpenAI or Anthropic provides, but focused on both the Chinese market and increasingly international developers.

Founders

The company was founded by former Microsoft executives:

Jiang Daxin (CEO)
Zhu Yibo
Jiao Binxing

Jiang Daxin previously served as a Microsoft Vice President and has a strong systems/software engineering background. (Wikipedia)

Funding

StepFun has attracted major Chinese investors, including:

Tencent
Qiming Venture Partners
Shanghai state-backed investment funds

According to recent reporting, the company has raised multiple large funding rounds and has been preparing for a Hong Kong IPO with valuations reported around US$10–12 billion. (Wikipedia)

Model Family

StepFun develops several model families.

Step 3.x

Their flagship LLM series.

Recent releases include:

Step 3
Step 3.5 Flash
Step 3.7 Flash

These models emphasize:

coding
agents
tool use
long context
multimodal reasoning

Their official platform describes Step 3.7 Flash as optimized for production agents with reliable tool orchestration, browser/terminal interaction, MCP compatibility, and multimodal understanding. (StepFun)

Step-Audio

One of the stronger open speech models from China.

Capabilities include:

speech recognition
speech synthesis
voice cloning
emotion control
multilingual speech

They also published the research paper and open-sourced parts of the system. (arXiv)

Step Image

Image generation and editing models intended to compete with systems like FLUX, GPT Image, and Midjourney for developer workflows. (StepFun)

Agent Models

A major strategic direction.

Research

Compared with many startups that mainly ship APIs, StepFun publishes a significant amount of research.

Examples include:

Step-3 system/model co-design
Step-Audio
Step-Prover
Step-Formalizer

Topics include:

efficient inference
theorem proving
speech models
RL
multimodal learning
reasoning
agent systems

(arXiv)

Open Source

They have begun open-sourcing several projects, including:

Step-Audio
StepDeepResearch
model checkpoints for some releases

Recent Step 3.7 Flash releases are available in formats including BF16, FP8, and GGUF for local inference. (Reddit)

Developer Platform

Their API platform provides:

Chat Completion APIs
Vision APIs
Image APIs
Speech APIs
Embeddings
Agent-oriented models
Subscription plans (“Step Plan”)

The platform is targeting developers building with tools such as Cursor, Claude Code, OpenClaw, and other agent frameworks. (StepFun)

Why StepFun Matters

From an engineering perspective, StepFun is interesting because it is focusing on several frontier areas simultaneously:

Production-grade agent models
Multimodal reasoning
Efficient MoE architectures
Model–system co-design for faster inference
Open-source speech technology
Long-context coding models

If you’re interested in LLM infrastructure, inference optimization, or agent engineering, StepFun is one of the Chinese labs worth following alongside DeepSeek and Moonshot.

References:

Back Donate