Join our daily and weekly newsletter for the latest updates and exclusive content on industry-leading AI coverage. learn more
Boston-based foundation model startup Liquid AI is spin-out from the Massachusetts Institute of Technology (MIT) and is trying to move the technology industry beyond its reliance on the most popular and large-scale language models (LLMS), such as Openai’s GPT series and Google’s Gemini family.
Yesterday, the company announced the Hyena Edge, a new convolution-based multi-hybrid model designed for smartphones and other edge devices, ahead of the International Conference on Learning Expression (ICLR) 2025.
One of the best events for machine learning research, the conference will be held this year in Vienna, Austria.
New convolution-based models promise faster and more memory-efficient AI at the edge
The Hyena Edge is designed to surpass the powerful transformer baseline in both computational efficiency and language model quality.
In real-world testing on a Samsung Galaxy S24 Ultra smartphone, the model provided lower latency, a smaller memory footprint and better benchmark results compared to a parameter-matched transformer++ model.
New architecture for the new era of edge AI
Unlike most small models designed for mobile deployments, including the SMOLLM2, PHI models and the Llama 3.2 1B, the Hyena Edge is away from traditional attention-rich designs. Instead, we strategically replace two-thirds of grouped query notes (GQA) operators with gate convolutions in the Hyena-Y family.
The new architecture is the result of the synthesis of a tailored architecture (STAR) framework for liquid AI, which automatically designs the model backbone using evolutionary algorithms, and was announced in December 2024.
Star explores a wide range of operator compositions rooted in mathematical theory of linear input variable systems to optimize multiple hardware-specific goals such as latency, memory usage, and quality.
Benchmarked directly on consumer hardware
To validate the real-world preparation of the Hyena Edge, Liquid AI ran tests directly on a Samsung Galaxy S24 Ultra smartphone.
The results show that Hyena Edge achieves up to 30% faster prefill and decoding latency compared to its transformer++ counterpart, increasing the speed advantage over longer sequence lengths.
Prefill latency with short sequence lengths exceeded the trans baseline. This is a key performance metric for responsive-on-device applications.
From a memory point of view, Hyena Edge placed it as a strong candidate for resource constraints as it did not consistently use LAM during inference across all tested sequence lengths.
Better than trance on language benchmarks
Hyena Edge was trained with 100 billion tokens and was evaluated across standard benchmarks for small language models such as Wikitext, Lambada, Piqa, Hellaswag, Winogrande, Arc-Easy, and Arc-Challenge.
In all benchmarks, Hyena Edge has matched or exceeded the performance of the GQA-Transformer++ model, improving notable improvements in Wikitext and Lambada’s confusion scores, as well as higher accuracy rates for Piqa, Hellaswag, and Winogrande.
These results suggest that the increased efficiency of the model does not sacrifice predictive quality.
Hyena Edge Evolution: View Performance and Operator Trends
For those looking for a deeper dive into the Hyena Edge development process, the recent video walkthrough offers an engaging visual summary of the evolution of the model.
This video highlights how key performance metrics, including Prefill LaTency, Decode Latency and Memory Consuption, improve the improvements to the architecture of consecutive generations.
It also provides a rare behind the scenes look at how the internal configuration of the hyena edge shifted during development. Viewers can see dynamic changes in the distribution of operator types, such as auto-joint (SA) mechanisms, various hyena variants, and Swiglu layers.
These shifts provide insight into architectural design principles that will help the model reach current levels of efficiency and accuracy.
By visualizing trade-offs and operator dynamics over time, this video provides valuable context for understanding the architectural breakthroughs underlying Hyena Edge’s performance.
Open Source Plan and a Broader Vision
Liquid AI said it plans to open source a range of Liquid Foundation models, including Hyena Edge, over the next few months. The company’s goal is to build competent and efficient general purpose AI systems that can scale from cloud data centers to personal edge devices.
The Hyena Edge debut highlights the growing potential for alternative architectures that challenge transformers in real-world settings. As mobile devices are increasingly expected to run sophisticated AI workloads natively, models like Hyena Edge can set new baselines that edge-optimized AI can achieve.
The Success of Hyena Edge – Both raw performance metrics and introductions to automated architectural design, liquid AI positions itself as one of the new players to see in the landscape of evolving AI models.
Source link
