StepFun

StepFun, formally Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd., is a Chinese artificial-intelligence company headquartered in Shanghai. Founded in April 2023 by former Microsoft employees, the company develops large foundation models intended to process and generate several forms of information, including text, images, video, and audio. Its work is centered on the Step family of models, spanning general-purpose language models, multimodal systems, image-generation tools, video generation, audio understanding, and visual reasoning. The company emerged during the rapid expansion of China's large-model sector and has been described by investors and industry observers as one of China's six so-called AI Tiger companies. Its investors have included Tencent, Qiming Venture Partners, and Shanghai State-owned Capital Investment. These relationships place StepFun within a wider ecosystem of Chinese technology companies, venture investors, state-linked capital, model developers, and domestic semiconductor manufacturers. StepFun publicly expanded its model portfolio at the World Artificial Intelligence Conference in July 2024. It introduced Step-2, described as a trillion-parameter large language model, together with Step-1.5V, a multimodal model, and Step-1X, an image-generation model. The releases indicated a strategy broader than text-only chatbots: the company has emphasized models capable of interpreting and producing content across multiple modalities and of supporting applications in areas such as intelligent software, enterprise services, content creation, and embodied or device-based systems. In 2025, StepFun added models aimed at global developer access and specialized reasoning. In February, it and Geely announced the open-sourcing of Step-Video-T2V, a text-to-video model, and Step-Audio, an audio-oriented multimodal model. In April, the company released Step-R1-V-Mini, a visual reasoning model designed for image interpretation and understanding. In July, it released Step 3 and announced an ecosystem alliance intended to improve the model's performance on Chinese domestic AI chips. StepFun's ecosystem strategy became particularly visible at the July 2025 World Artificial Intelligence Conference. The company announced a Model-Chip Ecosystem Innovation Alliance involving large-model developers and Chinese chip companies including Huawei, Biren Technology, Moore Threads, and Enflame. A separate Shanghai General Chamber of Commerce AI Committee was also established with participation from StepFun, SenseTime, MiniMax, MetaX, and Iluvatar CoreX. These initiatives reflect the company's effort to connect model development with local computing infrastructure and industry deployment. In February 2026, StepFun released Step 3.5 Flash, described as a mixture-of-experts model with 196 billion total parameters and 11 billion active parameters. The model was reported as being made available under the Apache 2.0 license, with tool-use capabilities and a 256,000-token context window. The same month, reports said the company was seeking an initial public offering on the Hong Kong Stock Exchange; this was a reported fund-raising or listing intention rather than evidence that the company had already become publicly listed. StepFun's positioning combines frontier-model research, open model releases, multimodal product development, and adaptation to China's domestic semiconductor ecosystem. Its public profile is therefore shaped both by the capabilities of its Step models and by its participation in partnerships and alliances intended to accelerate commercial and industrial adoption.

StepFun was established in April 2023 in Shanghai by former Microsoft employees. The company entered China's increasingly competitive foundation-model market with a focus on large language models and, from an early stage, multimodal systems capable of handling more than text alone. Its investors have included Tencent, Qiming Venture Partners, and Shanghai State-owned Capital Investment. The combination of private technology investment and state-linked capital reflects the importance of large-model development within China's broader artificial-intelligence and computing strategy. The company became known for the Step series of models. Its public product direction broadened substantially at the World Artificial Intelligence Conference in July 2024, where StepFun launched Step-2 alongside Step-1.5V and Step-1X. Step-2 was presented as a trillion-parameter large language model. Step-1.5V extended the portfolio into multimodal understanding, while Step-1X addressed image generation. Together, the releases showed that StepFun was pursuing a family of foundation models rather than a single general chatbot product. StepFun continued this expansion in 2025. In February, it worked with Geely on the open-sourcing of Step-Video-T2V and Step-Audio, making video-generation and audio-related capabilities available to developers beyond the company itself. In April, StepFun released Step-R1-V-Mini, a visual reasoning model designed to interpret images and support image-understanding tasks. In July, the company released Step 3 and connected the model's development to a domestic hardware strategy. At the July 2025 World Artificial Intelligence Conference, StepFun announced the Model-Chip Ecosystem Innovation Alliance. Participants included Chinese model developers and semiconductor companies such as Huawei, Biren Technology, Moore Threads, and Enflame. The stated purpose was to promote cooperation between model makers and chip manufacturers and to optimize models for domestic computing platforms. A second initiative, the Shanghai General Chamber of Commerce AI Committee, included StepFun together with SenseTime, MiniMax, MetaX, and Iluvatar CoreX. These activities positioned StepFun not only as a model developer but also as a participant in the organization of China's AI software-and-hardware ecosystem. In February 2026, StepFun released Step 3.5 Flash. It was described as a mixture-of-experts model with 196 billion total parameters and 11 billion active parameters. The model was reported to support tool use, a 256,000-token context window, and distribution under the Apache 2.0 open-source license. Also in February 2026, reports indicated that StepFun was seeking a Hong Kong Stock Exchange initial public offering. That report represented a potential future listing and did not establish that StepFun had already become a listed company. Across these stages, StepFun's development has followed several parallel tracks: scaling general-purpose models, adding multimodal capabilities, releasing specialized systems for image, video, audio, and reasoning tasks, making selected models available to external developers, and improving compatibility with China's domestic chip ecosystem. The company is commonly grouped by observers with China's leading private AI startups, sometimes under the informal label of the six AI Tigers.

1. 2026Step 3.5 Flash released

   StepFun released Step 3.5 Flash with mixture-of-experts architecture, tool-use support, a long context window, and an Apache 2.0 license.

   Source
2. 2025Open-source video and audio models announced

   StepFun and Geely announced open-source releases of Step-Video-T2V and Step-Audio.

   Source
3. 2025Step-R1-V-Mini released

   StepFun introduced a multimodal visual-reasoning model for image interpretation and understanding.

   Source
4. 2025Step 3 and model-chip alliance announced

   The company released Step 3 and announced cooperation among model developers and Chinese chip manufacturers.

   Source
5. 2024Step-2 and multimodal portfolio launched

   At WAIC, StepFun launched Step-2, Step-1.5V, and Step-1X, covering large-language, multimodal, and image-generation use cases.

   Source
6. 2023Company founded

   StepFun was founded in April 2023 in Shanghai by former Microsoft employees.

   Source

A Chinese foundation-model and AI infrastructure provider focused on multimodal intelligence, open model development, and compatibility with domestic AI chips.

• Step-1Large language model

  Step-1 is part of StepFun's foundational large-model family and forms the basis for the company's broader model-development program. Public descriptions identify it as a general large language model rather than a narrowly specialized system.

• Step-2Large language model2024

  Step-2 is a large language model launched at the July 2024 World Artificial Intelligence Conference. It was described as having a trillion parameters and represents StepFun's effort to build a large general-purpose foundation model for language and AI applications.

• Step-1.5VMultimodal model2024

  Step-1.5V is a multimodal model introduced with Step-2 in 2024. Its role is to extend model interaction beyond text by supporting visual or other mixed-input understanding.

• Step-1XImage-generation model2024

  Step-1X is StepFun's image-generation model announced at WAIC in 2024. It broadened the company's product family into generative visual content rather than limiting it to language and comprehension.

• Step-Video-T2VText-to-video model2025

  Step-Video-T2V is a text-to-video generation model developed by StepFun with Geely. The companies announced its open-source release in February 2025, making it available as part of the company's effort to engage external developers.

• Step-AudioAudio and multimodal model2025

  Step-Audio is an audio-oriented model released through StepFun's 2025 collaboration with Geely. It extends the Step portfolio into audio processing and multimodal applications and was announced for open-source access.

• Step-R1-V-MiniVisual reasoning model2025

  Step-R1-V-Mini is a multimodal reasoning model released in April 2025. It is designed for visual interpretation and image understanding, adding a reasoning-focused capability to StepFun's multimodal portfolio.

• Step 3Foundation model2025

  Step 3 is a later-generation StepFun foundation model released in July 2025. Its launch was associated with an ecosystem initiative intended to optimize model operation on Chinese domestic AI chips.

• Step 3.5 FlashMixture-of-experts foundation model2026

  Step 3.5 Flash was released in February 2026 as a mixture-of-experts model reported to contain 196 billion total parameters and 11 billion active parameters. It supports tool use and a 256,000-token context window and was made available under the Apache 2.0 license.

• 2025Global open-source model collaboration with Geely

  Global developer community

  StepFun and Geely announced the open-sourcing of Step-Video-T2V and Step-Audio, using developer access to encourage experimentation with video-generation and audio models.

  Outcome. The two models were announced for open-source access; broader commercial or adoption outcomes are not specified in the available material.

  Source
• 2025Model-Chip Ecosystem Innovation Alliance

  China

  Announced at WAIC, the alliance connected StepFun and other Chinese large-model developers with domestic chip manufacturers, including Huawei, Biren Technology, Moore Threads, and Enflame.

  Outcome. The initiative was intended to improve cooperation and model optimization across China's AI software and hardware ecosystem.

  Source
• 2025Shanghai General Chamber of Commerce AI Committee

  Shanghai · China

  StepFun joined a Shanghai AI committee established alongside SenseTime, MiniMax, MetaX, and Iluvatar CoreX.

  Outcome. The committee created an industry-cooperation forum; specific commercial results are not stated in the available material.

  Source

• 2026Release Step 3.5 Flash under Apache 2.0Product launch

  StepFun continued its progression toward larger and more capable foundation models while maintaining an open model distribution strategy.

  What changed. The company released Step 3.5 Flash as a mixture-of-experts model with tool use, a 256,000-token context window, and an Apache 2.0 license.

  Aftermath. The release expanded StepFun's open model portfolio; longer-term adoption and commercial impact are not specified.

  Source
• 2026Reported consideration of a Hong Kong IPOOther

  Reports in February 2026 said StepFun was seeking an initial public offering on the Hong Kong Stock Exchange.

  What changed. The reported plan was to pursue a potential Hong Kong listing; no completed listing or offering details are established in the available material.

  Aftermath. StepFun remains classified here as a private company because the source describes a prospective transaction rather than a completed IPO.

  Source
• 2025Pursue domestic model-chip optimizationStrategy

  The release of Step 3 coincided with StepFun's participation in an alliance linking Chinese model developers and AI-chip manufacturers.

  What changed. StepFun supported the Model-Chip Ecosystem Innovation Alliance to encourage optimization of its models for domestic Chinese chips.

  Aftermath. The initiative strengthened StepFun's public positioning as a participant in China's domestic AI computing ecosystem; measurable business effects are not specified.

  Source
• 2025Expand external developer access through open-source releasesStrategy

  StepFun sought to extend its model portfolio beyond proprietary general language systems into video and audio applications.

  What changed. Together with Geely, the company announced open-source access to Step-Video-T2V and Step-Audio.

  Aftermath. The releases made selected Step models available to global developers, although the available material does not quantify adoption or revenue.

  Source

• 2026StepFun reported to be considering a Hong Kong IPO

  Reports said StepFun was seeking an initial public offering on the Hong Kong Stock Exchange. The report described a prospective listing, not a completed flotation.

  Other
  Source
• 2026StepFun releases Step 3.5 Flash

  StepFun released Step 3.5 Flash, a mixture-of-experts model reported to have 196 billion total parameters, 11 billion active parameters, tool-use support, a 256,000-token context window, and an Apache 2.0 license.

  Product launchProduct generation
  Source
• 2025StepFun and Geely announce open-source video and audio models

  StepFun and Geely announced the open-sourcing of Step-Video-T2V and Step-Audio for global developers.

  Product launch
  Source
• 2025StepFun releases Step-R1-V-Mini

  The company released Step-R1-V-Mini, a multimodal reasoning model intended for visual interpretation and image understanding.

  Product launch
  Source
• 2025StepFun releases Step 3 and announces model-chip alliance

  StepFun released Step 3 and announced an alliance with Chinese AI-model developers and chip manufacturers to improve model optimization on domestic processors.

  Product launch
  Source
• 2024StepFun launches Step-2 and multimodal models at WAIC

  At the World Artificial Intelligence Conference, StepFun introduced Step-2, Step-1.5V, and Step-1X, expanding its portfolio from general language modeling into multimodal understanding and image generation.

  Product launchProduct generation
  Source

1. StepFun
2. StepFun official website