Jiang Daxin

Overview

Jiang Daxin is the founder and chief executive officer of StepFun, a Chinese company working on general-purpose large models. He holds a PhD, and is described as a former Microsoft global vice president and chief scientist of the Microsoft Asia Internet Engineering Institute^[1].

Research and engineering background

His research and engineering work spans machine learning, data mining, natural language processing and bioinformatics^[2][1]. Jiemian reported in June 2024 that he spent 16 years at Microsoft^[4].

Those four areas sit close to the work of building large models. Machine learning supplies the methods by which a model's parameters are fitted to data, typically by adjusting them in the direction that reduces error. Natural language processing supplies the tasks — translation, question answering, summarisation — and the habits of measuring systems against test sets. Data mining and bioinformatics are fields that have long dealt with large, noisy datasets where the useful signal has to be extracted statistically rather than by hand-written rules. A career spent across them is a career spent on the parts of the problem that a model company has to solve at scale: how to learn from data, and how to tell whether the learning worked.

Founding StepFun

In March 2023 he was promoted to Microsoft vice president. About a month later he registered StepFun; Jiemian reported the registered entity as Shanghai StepFun Intelligent Technology Co., Ltd.^[2][4]

At the public debut: two claims

StepFun made its public debut in March 2024. At that debut he argued that scaling laws set the ceiling on model intelligence, and that unifying multimodal understanding and generation is necessary for AGI^[1]. Jiemian reported the same position as StepFun's view — that unifying multimodal understanding and generation is a necessary path to AGI^[4].

Scaling laws are an empirical regularity: as a model's size, the amount of data it is trained on and the compute spent on training grow, its error on many tasks falls in a fairly predictable way. To say that scaling laws set the ceiling on intelligence is to say that the outer limit of what a model can reach is decided above all by how far that curve is pushed, rather than by cleverness added on top of a small model.

The second claim is about what a model should be able to handle. Multimodal understanding means a model takes in text, images, audio or video and extracts meaning from them; multimodal generation means it produces such content. These were long handled by separate systems, one for reading and one for making. Unifying them puts both abilities inside one model, so that what the model learns about the structure of images or speech can inform what it writes, and what it writes can be rendered back into images or audio. His stated reason for wanting this is that AGI, on his account, requires it — not that it is merely convenient.

Management change

In January 2026, Megvii founder Yin Qi became StepFun's chairman. Jiang Daxin remained chief executive officer after that appointment^[3].

Position in the record

His public statements form a compact statement of where he thinks the limits and the direction of large-model work lie: the ceiling is set by scale, and the direction is toward a single model that both understands and generates across media. Both claims are attached to StepFun's launch and to the company's own account of its approach^[1][4]. His career detail — a PhD, a long Microsoft tenure, senior research management, then founding and running a model company — is the pattern that several Chinese large-model startups share, and it is the reason his move from a large technology company into a startup was followed closely^[2][4].