Liquid AI Unveils Hyena Edge Model for Faster Efficient AI on Smartphones
Liquid AI, a startup from MIT, introduced Hyena Edge, a new convolution-based AI model designed to surpass Transformer-based models in efficiency and quality on edge devices. Tested on Samsung Galaxy S24 Ultra, Hyena Edge delivers up to 30% faster latency and lower memory use while matching or exceeding language benchmark performance. This innovation signals a shift toward more efficient AI architectures optimized for mobile and edge computing.
Liquid AI, a Boston-based startup spun out of the Massachusetts Institute of Technology (MIT), is pioneering a new frontier in artificial intelligence by moving beyond the dominant Transformer architecture. Their latest innovation, Hyena Edge, is a convolution-based, multi-hybrid AI model designed specifically for smartphones and edge devices, promising significant improvements in speed, memory efficiency, and language understanding.
Unlike traditional Transformer models that rely heavily on attention mechanisms, Hyena Edge replaces two-thirds of grouped-query attention (GQA) operators with gated convolutions from the Hyena-Y family. This architectural shift is the result of Liquid AI’s proprietary Synthesis of Tailored Architectures (STAR) framework, which uses evolutionary algorithms to optimize model design for hardware-specific objectives such as latency, memory usage, and predictive quality.
Real-world testing on the Samsung Galaxy S24 Ultra smartphone demonstrated Hyena Edge’s superior performance, achieving up to 30% faster prefill and decode latencies compared to a parameter-matched Transformer++ model. Additionally, it maintained a smaller memory footprint across all tested sequence lengths, making it highly suitable for resource-constrained edge environments where responsiveness and efficiency are critical.
Hyena Edge was trained on an extensive dataset of 100 billion tokens and evaluated against standard benchmarks for small language models, including Wikitext, Lambada, PiQA, HellaSwag, Winogrande, and ARC challenges. It consistently matched or outperformed Transformer-based models, delivering better perplexity scores and higher accuracy rates without sacrificing efficiency. This balance of speed, memory use, and quality marks a significant advancement in edge AI capabilities.
The development journey of Hyena Edge is documented in a detailed video walkthrough that showcases the evolution of its architecture. Viewers gain insight into how performance metrics such as latency and memory consumption improved through successive iterations, alongside shifts in operator composition including Self-Attention, Hyena variants, and SwiGLU layers. This transparency highlights the innovative design principles driving the model’s success.
Looking ahead, Liquid AI plans to open-source Hyena Edge and other foundation models to foster broader adoption and innovation. Their vision is to create scalable AI systems that perform efficiently across the spectrum from cloud datacenters to personal edge devices. As mobile devices increasingly demand sophisticated AI capabilities natively, Hyena Edge exemplifies the potential for alternative architectures to redefine the standards of edge-optimized AI.
Hyena Edge’s breakthrough performance and automated architecture design position Liquid AI as a key emerging player in the AI model landscape. Their approach not only challenges the Transformer dominance but also opens new pathways for deploying powerful, efficient AI on everyday devices, enabling real-time, on-device intelligence with reduced latency and resource demands.
AI Tools Built for Agencies That Move Fast.
Explore how QuarkyByte’s AI insights can help your business leverage cutting-edge models like Hyena Edge for mobile and edge AI applications. Discover tailored strategies to optimize AI deployment on smartphones and edge devices, boosting performance while reducing resource consumption. Partner with QuarkyByte to stay ahead in the evolving AI landscape.