Skip to main content

Stepfun

Category: Foundation Models / LLMs

Stepfun (阶跃星辰) is a leading Chinese AI startup dedicated to developing trillion-parameter multimodal foundation models and AGI infrastructure, known as one of China's 'AI Six Little Tigers.' Stepfun was founded in 2023. The company is led by Daxin Jiang. Based in Shanghai, China. Team size: 500+. Total funding raised: $3,800.0M. Latest round: Series B+ ($704.0M, Jan 2026). Key investors include Shanghai State-Owned Capital Investment, Tencent, Meituan, Qiming Venture Partners, 5Y Capital, China Life Equity Investment, Pudong Venture Capital, Alibaba, Huaqin Technology, Longcheer Technology, OmniVision (OmniVision), ZTE, Hong Kong Investment Corporation (HKIC).

Founded
2023
Headquarters
Shanghai, China
Team size
500+
Total funding
$3,800.0M

Value proposition

Delivers native multimodal intelligence at scale, combining massive parameter counts (trillion-scale) with high inference efficiency for real-world industry applications.

Products and solutions

Step-1 & Step-1V: 100B-parameter language and multimodal models, Step-2: 1.3T-parameter Mixture of Experts (MoE) foundation model, Step-3 Series: 321B-parameter MoE model with native multimodal reasoning, Step-3.5-Flash: 196B-parameter MoE model (11B active) released Feb 2026 under Apache 2.0, Step-Video-T2V & Step-Audio: Specialized open-source models for video generation and audio interaction, Step-Audio 2.5 TTS: Text-to-speech model released Apr 2026, Yuewen (跃问): Consumer-facing AI personal assistant and chatbot, Maopao Ya (冒泡鸭): AI-powered character role-playing and companion app, Stepfun Open Platform: API services for developers and enterprise model customization

Unique value

Stepfun distinguishes itself by being one of the first in China to successfully train a trillion-parameter MoE model while maintaining 'Native Multimodality'—integrating vision, audio, and text at the architectural level rather than as separate modules.

Target customer

Enterprise clients (Smartphone OEMs, Automotive, Finance), independent developers, and individual consumers.

Industries served

Consumer Electronics (Smartphones/IoT), Automotive (Intelligent Cockpits), Financial Services, Retail & E-commerce, Digital Entertainment & Content Creation

Technology advantage

Leverages a proprietary MFA (Attention mechanism) to reduce KV cache overhead and power-efficient MoE (Mixture of Experts) architectures. The company demonstrates a 'Scaling Law' leadership, with models like Step-3 achieving 300% higher inference efficiency on domestic chips compared to competitors like DeepSeek-R1.

How they differentiate

Stepfun differentiates through a commitment to 'Native Multimodality' and the 'Scaling Law,' being one of the first in China to successfully train a trillion-parameter Mixture of Experts (MoE) model (Step-2). Unlike competitors focusing solely on text or context length, Stepfun integrates vision, audio, and text at the architectural level and targets 'AI+Terminal' applications for automobiles and smartphones.

Main competitors

Moonshot AI (Kimi), Zhipu AI, MiniMax, DeepSeek

Key partnerships

Geely Auto: Jointly open-sourced Step-Video and Step-Audio models for automotive applications., OPPO & Honor: Integration of Step-series models into next-generation flagship smartphones., Shanghai State-Owned Capital Investment: Strategic backing and resource alignment., Mo-Chip Ecosystem Innovation Alliance: Collaboration with domestic chip manufacturers to optimize model-hardware synergy.

Notable customers

Geely Auto, OPPO, Honor, Huanrui Century

Major milestones

Founded in April 2023 by former Microsoft VP Daxin Jiang, Launched Step-2, China's first trillion-parameter MoE model in March 2024, Reached unicorn status with $2B valuation in June 2024, Announced strategic 'AI+Terminal' partnership with Geely Auto in Feb 2025, Original Megvii CEO Yin Qi joined as Chairman in Jan 2026, Closed record-breaking 5 billion RMB B+ round in Jan 2026, Released Step-3.5-Flash open-source model under Apache 2.0 in Feb 2026, Completed shareholding reform and dismantled red-chip structure for HK IPO in Apr 2026, Closed ~$2.5B Pre-IPO funding round with strategic industrial investors in May 2026

Growth metrics

API call volume increased 45x within a 6-month period in 2024-2025; valuation reached $2 billion by mid-2024 before the massive B+ round.

Market positioning

Tier-1 Chinese Generative AI unicorn; one of the 'AI Six Little Tigers' leading China's AGI race.

Geographic focus

Mainland China (Primary), Global (Open-source and API services)

Patents and IP

Holds proprietary intellectual property in MFA Attention mechanisms, high-efficiency vision encoders (reducing visual tokens by 16x), and GUI Agent technologies (GELab-Zero).

About Daxin Jiang

Daxin Jiang (姜大昕) is a distinguished AI leader and former Corporate Vice President at Microsoft. He spent over 16 years at Microsoft, most recently as the Chief Scientist at Microsoft Asia Internet Engineering Research Institute (STCA). He led the engineering and research for Bing search, Cortana, and NLP systems for Microsoft 365 and Azure Cognitive Services. He is a globally recognized expert in data mining and NLP with over 100 research papers published.

Official website: