iFold, Apple's new AI achievement
Whoa! Why is Apple venturing into the cross - border AI model field?
Apple has released a protein folding model called SimpleFold based on flow matching, which netizens jokingly refer to as “iFold”.
SimpleFold doesn't have fancy and exclusive module designs. It relies on the general Transformer module combined with the flow - matching generation paradigm. The 3B parameter version has matched the performance of Google's AlphaFold2, the leading model in this field.
It seems that Apple's cross - border move here is about simplifying complexity.
It runs smoothly on a MacBook Pro
Let's first understand what protein folding is all about.
The core is to fold a “string” of amino acids into a specific 3D shape so that the protein can function.
A protein folding model predicts the three - dimensional spatial conformation of a protein from its primary amino acid sequence.
Previously, the most powerful models, such as Google's AlphaFold2, although achieving breakthroughs, used many complex and exclusive designs.
For example, they need to analyze the sequences of a large number of similar proteins, rely on multiple sequence alignment (MSA) to build evolutionary information, optimize spatial constraints through triangular attention, and require super - computing power during inference, which is unaffordable for ordinary laboratories.
However, this “iFold” solves this problem with a general AI framework.
SimpleFold uses a multi - layer Transformer encoder as the core backbone in its architecture. It only adapts to protein sequence features through adaptive layer normalization, which is equivalent to using a “general toolbox” to solve problems in a specific field.
The core innovation lies in the introduction of flow - matching generation technology.
Different from the step - by - step denoising of diffusion models, flow matching realizes one - step generation of atomic coordinates by learning a smooth mapping from a random noise distribution to a protein conformation distribution.
During the training phase, the team built a mixed dataset containing 9 million data entries and trained multi - scale models with parameters ranging from 100M to 3B. Among them, SimpleFold - 3B achieved 95% of the performance of AlphaFold2 in the CAMEO22 benchmark test.
It outperformed the similar flow - matching model ESMFold on the high - difficulty CASP14 test set.
It's also worth mentioning its efficiency. On a MacBook Pro equipped with an M2 Max chip, the inference time for processing a 512 - residue sequence is only two or three minutes, far exceeding the hours - long time consumption of traditional models.
The research team
The first author of this research, Yuyang Wang, graduated from Tongji University with a bachelor's degree. Then he went to Carnegie Mellon University in the United States for further studies and successively obtained a master's degree in mechanical engineering, a master's degree in machine learning, and a doctorate in mechanical engineering. His long - term studies have laid a solid foundation for his research in related fields.
He has an internship experience in Momenta, engaged in the research and development of reinforcement learning. He also served as an AI/ML Resident at Apple, focusing on diffusion model research, and later became a machine learning researcher at Apple.
The corresponding author is the Chinese machine learning engineer Jiarui Lu, who graduated from Tsinghua University with a bachelor's degree. During his studies, he also served as a research assistant in Professor Jun Zhu's laboratory.
Subsequently, Lu obtained a master's degree in machine learning from Carnegie Mellon University and joined Apple in 2020 after graduation.
He once led the development of a benchmark for the tool - calling ability of large models, ToolSandbox, an open - source achievement of Apple.
Friends who are interested in this “iFold” and want to dig into the technical details can click the link at the end of the article.
Paper link: https://arxiv.org/abs/2509.18480 Code link: https://github.com/apple/ml - simplefold
Reference link: [1]https://x.com/iScienceLuvr/status/1970787581248905454
This article is from the WeChat public account “QbitAI”. The author focuses on cutting - edge technology. 36Kr is authorized to publish it.