CUPERTINO, Calif.--()--Aizip, Inc. in partnership with SoftBank Corp., announced the release of customized Small Language Model (SLM) and Retrieval Augmented Generation (RAG) solutions for enterprise applications. The system operates locally on mobile devices or on-premises servers, addressing the pressing concerns of enterprise data safety. Fine-tuned with domain specific data, these SLMs can address unique enterprise tasks with comparable accuracy to 100x larger cloud LLMs.

This announcement comes as many companies have implemented bans on cloud-based LLMs for business-related tasks, citing concerns about leaks of private information. A recent survey from Menlo Ventures found that 21% of failed AI pilots were due to data-privacy hurdles. Fine-tuned SLMs are an alternative to general LLMs, offering security and privacy of information along with the benefits of AI-enabled work.

Many enterprises have explored using open-source Small Language Models that run locally and efficiently on-premises as replacements for cloud-based LLMs. However, these models have generally turned out to be too unreliable to meet production-level performance requirements.

Aizip and SoftBank address this challenge with innovative new solutions. The joint team developed and delivered a customized SLM integrated into a RAG system, running locally as a mobile application. When applied to SoftBank’s internal documents the application offers employees a Q&A platform that generates accurate, up-to-date response, running entirely on an iPhone 14. Testing shows that the system satisfactorily addresses 97% of employee questions, and response quality measured on a dataset of 22,000 questions matched responses generated by a GPT-4o-based RAG system.

The key to this unique reliability is Aizip’s SLM-customization pipeline, which includes high-quality data generation, fine-tuning, and multifaceted evaluation. “While there are an increasing number of open-source SLMs available today, off-the-shelf models often fall short of the production accuracy required for enterprise tasks,” noted Aizip SLM-development lead Weier Wan. Aizip’s SLM-customization pipeline is run iteratively until target accuracy is achieved, so enterprises no longer need to choose between privacy and performance.

Although the first product was developed for the iPhone, Aizip’s flexible SLM-deployment tools allow SLMs and RAG systems to run on a variety of edge platforms, including on-premises servers, PCs, and even IoT devices like microprocessors (MPUs), as announced in June 2024.

Looking ahead, the joint team aims to leverage SoftBank’s rich expertise in enterprise services and Aizip’s cutting-edge AI technology to expand customized SLM solutions to a broader range of applications and customers. “Privacy-critical and offline-required use cases can benefit greatly from Aizip’s on-device AI technology,” described SoftBank vice president Katsuya Kitazawa, head of the Information Technology & Architect Division. “Whether assisting flight attendants on airplanes or supporting field workers in remote locations, we’re excited to collaborate with Aizip to bring this innovation to more users and businesses.”

Aizip remains committed to delivering production-grade AI solutions for a variety of on-device applications. With extensive experience in developing robust and efficient AI models across Aizip Intelligent Audio (AIA), Vision (AIV), and Time-Series (AIT) product lines, Aizip continues to pioneer SLM technology, focused on improved accuracy, reliability, speed, and development efficiency. For additional information, please contact info@aizip.ai.

About Aizip, Inc.

Situated in the heart of Silicon Valley, Aizip, Inc. specializes in developing superior AI models tailored for endpoint and edge-device applications. Aizip stands apart for its exemplary model performance, swift deployment, and remarkable return on investment. These models are versatile, catering to a spectrum of intelligent, automated, and interconnected solutions. Discover more at www.aizip.ai.