Starland.AI
The first GenAI Model-as-a-Service (MaaS) DePIN network, capable of running large multimodal models using any type of computing devices.
About
The problem it solves Our team, with its rich background in distributed systems, cloud computing, cloud-native technologies, blockchain, and AI PaaS platform development, is a seasoned startup backed by IDG Capital and Web3OG. We've leveraged cloud-native distributed system technology to fine-tune and deploy leading open-source large language models. These are integrated with advanced systems such as TTS voice recognition, vector databases, stable diffusion image generators, and music creation models, forming a robust multimodal MaaS (Model-as-a-Service) system. This platform offers a diverse range of applications including dialogue, imagery, music, sound, and digital avatars, accessible via APIs and SDKs. Our Web3 development partners simply need to define their interaction per second (TPS) requirements and desired model outcomes to utilize these models effortlessly. Additionally, our service seamlessly integrates with any blockchain system, supporting features like NFTs, smart contracts, and inscriptions. We employ cutting-edge technologies such as tensor parallelism, ollama, memory optimization, enhanced inference, fine-tuning, and edge computing, along with a cloud-native and container-based computational power management platform. This setup effectively utilizes underused computational resources, such as NVIDIA GPUs (models 3080 and 3090) constrained by Ethereum's POW limitations, and achieves optimal performance even on standard CPUs with minimal memory. Our capabilities ensure model performance on par with leading AI solutions like c.ai and GPT at significantly reduced costs. We have demonstrated this with our digital persona application, starland.ai, developed on the DePIN computational network. It matches GPT’s response speed and supports comprehensive multimodal capabilities while slashing computational costs by 90%. This dramatic cost reduction has instilled confidence among Web3 developers and projects, drawing interest from dozens of potential clients.
How it's made
Challenges we ran into The integration of various modalities such as text, images, sound, and databases within distributed cloud-native systems adds layers of complexity to the deployment and management of GenAI models. Moreover, the diverse range of computing power, spanning from high-performance server capabilities like H100 and A100 to consumer-grade hardware such as 4090, 3090, and 3080 GPUs, poses challenges in unified management and resource allocation. Optimizing the performance of large GenAI models on low-end computing resources, including integrated graphics and CPUs, is a formidable task. The inefficient utilization of these resources often leads to underperformance and idle capacity, hindering the scalability of GenAI applications. Despite the impressive computational capabilities of current DePIN networks, there is a notable gap in supporting the development, deployment, and maintenance of generative AI applications. This lack of focus on practical applications limits the real-world usage and adoption of GenAI technology. The advent of distributed and cloud-native technologies has paved the way for innovative solutions to address the challenges encountered in GenAI models. Leveraging these technologies, novel approaches have been developed to fine-tune and infer large multimodal models using low-end computing resources effectively. Real-time generation of diverse content such as images, sound, conversations, and music on DePIN devices is now achievable, rivaling the capabilities of traditional high-performance systems. The introduction of Large Model as a Service (MaaS) marks a significant milestone in the evolution of GenAI products. MaaS offers elastic and highly available large model applications, coupled with seamless integration capabilities with low-end computing DePIN technology. This enables developers to harness the power of GenAI in their applications without the burden of managing complex infrastructure, thereby democratizing access to advanced AI capabilities.