StepFun — Free and Fast AI Models from China

The Multimodal AI Model That Launched from Shanghai

In the crowded Chinese AI landscape, StepFun chose a different path: specializing in multimodal understanding from day one — not just text, but text, image, sound, and video together. This decision made it one of the “Six Little Tigers” of Chinese AI, and enabled it to raise $719 million in January 2026, exceeding the IPO proceeds of its larger competitors at the same time.

Founding Story: A Microsoft VP Decides to Create Something Special

StepFun was founded in April 2023 by Dr. Jiang Daxin, former Vice President of Microsoft and former head of Microsoft’s Asia Software Technology Center. He worked at Microsoft for 16 years, contributing to the development of Bing and Microsoft 365 products.

His motivation was clear: when OpenAI launched ChatGPT in November 2022, Jiang decided he could do something similar or better. He recruited colleagues from Microsoft: Zhu Yibo (who was leading AI infrastructure at ByteDance) and researcher Zhang Xiangyu.

The name “StepFun” is inspired by the “Step Function” mathematical function — a function that changes suddenly, not gradually — referring to the company’s vision of discontinuous progress in AI.

Their slogan, placed on the wall of their Shanghai office: “From the unique to the multimodal, to embodied intelligence, and finally AGI.”

Main Products and Models

Step-1 and Step-1.5V

The company’s first text model, followed by Step-1.5V which added image and text understanding together.

Step-1X (Image Model, 2024)

A specialized image generation model enabling commercial production.

Step-2 (July 2024)

The first Chinese language model to exceed one trillion parameters. Officially launched at the 2024 World AI Conference in Shanghai.

Step-R1-V-Mini (April 2025)

A multimodal reasoning model specialized for visual interpretation and image understanding.

Step-3 (July 2025)

The third generation with 321 billion total parameters and 38 billion active. Introduced two main architectural innovations:

• Multi-Matrix Factorization Attention (MFA): Reduces KV-cache memory consumption to ~22% compared to DeepSeek V3.
• Attention-FFN Disaggregation (AFD): Separates attention layers and feed-forward networks into specialized subsystems for better device efficiency.

Step 3.5 Flash (February 2026)

An open-source model (Apache 2.0) with 196 billion total parameters and 11 billion active:

• Context window of 262,000 tokens.
• Speed of 100-300 tokens/second.
• 99.8% on AIME 2025 math benchmark, 98.0% on HMMT 2025.

Yuewen (Consumer Interface)

An AI assistant app for ordinary users available on iOS, Android, and web, employing Step’s internal models.

Agent OS (in partnership with Geely)

An intelligent car operating system integrating StepFun’s multimodal models and voice recognition into vehicle interfaces. First launched in Geely Galaxy M9.

Achievements and Numbers

Funding tells the story of a company accelerating on the right track:

• Series A (2023-2024): Hundreds of millions of yuan from Tencent, Qiming Venture Partners, and 5Y Capital, at a $2 billion valuation.
• Series B+ (January 2026): Over 5 billion yuan (~$719 million) — the largest AI funding round in Asia for that quarter. Investors: government institutions (Shanghai SSCI, China Life Private Equity, Pudong Venture Capital) alongside Tencent, Qiming, and 5Y Capital.

Notable Comparison: The funding amount ($719 million) exceeded the IPO proceeds of competitors Zhipu AI ($558 million) and MiniMax in Hong Kong during the same week.

Commercial Expansion:

• Over 42 million devices shipped containing StepFun models.
• ~20 million daily active users through partnerships.
• 170% quarterly growth in edge API calls for three consecutive quarters by end of 2025.
• Partners represent ~60% of major Chinese smartphone makers (OPPO, Honor, ZTE).

Automotive: The company targets exceeding one million integrated vehicles by end of 2026.

Competition and Challenges

StepFun is part of a group media-dubbed the “Six Little Tigers” of Chinese AI (alongside Zhipu AI, MiniMax, Moonshot, Baichuan, and Zhizhun). In 2025-2026, the paths of these companies began to diverge:

• Zhipu AI and MiniMax: Chose IPOs in Hong Kong.
• Moonshot: Massive private funding rounds.
• StepFun: Stayed away from media hype and focused on commercial partnerships with Geely and OPPO.

Core Competitive Advantage: Edge Devices. While most AI companies compete on servers and cloud, StepFun ships its models inside phones and cars — a far less crowded competitive environment.

Major Challenge: US restrictions on exporting advanced chips to China hinder access to Nvidia H100/H200. The company relies on Huawei Ascend, Moore Threads, and other local chips, which are less efficient.

Future Vision 2026–2027

Hong Kong IPO: In February 2026, Bloomberg reported that StepFun is exploring an IPO that could raise ~$500 million. The path of competitors who went public (Zhipu and MiniMax) looks encouraging, especially with a wave of enthusiastic investor buying of Chinese AI stocks.

One Million Cars: The target for end of 2026 is exceeding one million vehicles integrated with Agent OS — if achieved, it would be a major step in connecting AI to physical infrastructure.

Robotics: The announcement of partnership with robotics company Agibot in March 2025 indicates expansion toward “Embodied Intelligence” according to the company’s original vision.

Analytical Conclusion

StepFun possesses what many competitors lack: a tangible commercial execution strategy based on partnerships with device and car makers, not just cloud contracts. 42 million devices shipped with its models is real market position, not just user registrations.

The company prefers calm over hype: “We are not capable enough in some areas and may miss some windows, but time will prove that consistency is a real advantage” — said one of its officials to KR Asia. If the Chinese AI market is moving toward localization in devices and cars, StepFun is in a strong position.