A small team of AI researchers from Adobe Inc., in collaboration with colleagues from Auburn University and Georgia Tech, has developed a compact language model (SLM) they claim can operate entirely on a smartphone without relying on cloud services. The team has released a paper detailing their innovation, named SlimLM, on the arXiv preprint server. As large language model (LLM) technology evolves, researchers worldwide continue to explore ways to enhance its capabilities. This new development marks a significant step in enabling document processing to be done locally, without needing cloud connectivity.
With the growing popularity of LLMs like ChatGPT, privacy concerns have become more prevalent. It’s not just individuals who are worried; businesses of all sizes are adopting AI tools for various operations, many of which require strong privacy protections.
The issue with current LLMs is that much of their processing and storage happens on cloud servers, which are vulnerable to hacking. Experts have pointed out that a potential solution is to run smaller language models (SLMs) locally, eliminating the need for cloud access and addressing privacy concerns.
Major tech companies such as Google, Apple, and Meta have been working toward this goal, developing apps that can function without cloud support. However, these solutions have not yet been widely adopted. SlimLM, according to its creators, is the first to bridge that gap by being available for real-world use, with plans to launch soon.
The team acknowledges that SlimLM’s ability to run locally is due to its focused design—it is not a general-purpose tool or chatbot. Instead, it is specialized for document-related tasks, such as summarization and answering specific questions about text. This narrow focus allows the model to operate with fewer parameters, with the smallest version running on just 125 million parameters, reducing the computational load on smartphones.
The researchers believe their app represents a significant move toward more localized AI applications, offering enhanced privacy for a variety of use cases.
By shadjava