Be part of our each day and weekly newsletters for the most recent updates and unique content material on industry-leading AI protection. Be taught Extra
Liquid AI, the Boston-based basis mannequin startup spun out of the Massachusetts Institute of Know-how (MIT), is in search of to maneuver the tech {industry} past its reliance on the Transformer structure underpinning hottest massive language fashions (LLMs) resembling OpenAI’s GPT collection and Google’s Gemini household.
Yesterday, the corporate introduced “Hyena Edge,” a brand new convolution-based, multi-hybrid mannequin designed for smartphones and different edge units prematurely of the Worldwide Convention on Studying Representations (ICLR) 2025.
The convention, one of many premier occasions for machine studying analysis, is going down this 12 months in Vienna, Austria.
New convolution-based mannequin guarantees sooner, extra memory-efficient AI on the edge
Hyena Edge is engineered to outperform robust Transformer baselines on each computational effectivity and language mannequin high quality.
In real-world exams on a Samsung Galaxy S24 Extremely smartphone, the mannequin delivered decrease latency, smaller reminiscence footprint, and higher benchmark outcomes in comparison with a parameter-matched Transformer++ mannequin.
A brand new structure for a brand new period of edge AI
In contrast to most small fashions designed for cell deployment — together with SmolLM2, the Phi fashions, and Llama 3.2 1B — Hyena Edge steps away from conventional attention-heavy designs. As a substitute, it strategically replaces two-thirds of grouped-query consideration (GQA) operators with gated convolutions from the Hyena-Y household.
The brand new structure is the results of Liquid AI’s Synthesis of Tailor-made Architectures (STAR) framework, which makes use of evolutionary algorithms to routinely design mannequin backbones and was introduced again in December 2024.
STAR explores a variety of operator compositions, rooted within the mathematical principle of linear input-varying programs, to optimize for a number of hardware-specific goals like latency, reminiscence utilization, and high quality.
Benchmarked immediately on shopper {hardware}
To validate Hyena Edge’s real-world readiness, Liquid AI ran exams immediately on the Samsung Galaxy S24 Extremely smartphone.
Outcomes present that Hyena Edge achieved as much as 30% sooner prefill and decode latencies in comparison with its Transformer++ counterpart, with velocity benefits growing at longer sequence lengths.
Prefill latencies at quick sequence lengths additionally outpaced the Transformer baseline — a vital efficiency metric for responsive on-device purposes.
By way of reminiscence, Hyena Edge constantly used much less RAM throughout inference throughout all examined sequence lengths, positioning it as a robust candidate for environments with tight useful resource constraints.
Outperforming Transformers on language benchmarks
Hyena Edge was educated on 100 billion tokens and evaluated throughout normal benchmarks for small language fashions, together with Wikitext, Lambada, PiQA, HellaSwag, Winogrande, ARC-easy, and ARC-challenge.
On each benchmark, Hyena Edge both matched or exceeded the efficiency of the GQA-Transformer++ mannequin, with noticeable enhancements in perplexity scores on Wikitext and Lambada, and better accuracy charges on PiQA, HellaSwag, and Winogrande.
These outcomes recommend that the mannequin’s effectivity features don’t come at the price of predictive high quality — a standard tradeoff for a lot of edge-optimized architectures.
Hyena Edge Evolution: A take a look at efficiency and operator developments
For these in search of a deeper dive into Hyena Edge’s growth course of, a latest video walkthrough offers a compelling visible abstract of the mannequin’s evolution.
The video highlights how key efficiency metrics — together with prefill latency, decode latency, and reminiscence consumption — improved over successive generations of structure refinement.
It additionally presents a uncommon behind-the-scenes take a look at how the interior composition of Hyena Edge shifted throughout growth. Viewers can see dynamic modifications within the distribution of operator sorts, resembling Self-Consideration (SA) mechanisms, varied Hyena variants, and SwiGLU layers.
These shifts provide perception into the architectural design rules that helped the mannequin attain its present degree of effectivity and accuracy.
By visualizing the trade-offs and operator dynamics over time, the video offers useful context for understanding the architectural breakthroughs underlying Hyena Edge’s efficiency.
Open-source plans and a broader imaginative and prescient
Liquid AI mentioned it plans to open-source a collection of Liquid basis fashions, together with Hyena Edge, over the approaching months. The corporate’s purpose is to construct succesful and environment friendly general-purpose AI programs that may scale from cloud datacenters down to private edge units.
The debut of Hyena Edge additionally highlights the rising potential for various architectures to problem Transformers in sensible settings. With cell units more and more anticipated to run refined AI workloads natively, fashions like Hyena Edge may set a brand new baseline for what edge-optimized AI can obtain.
Hyena Edge’s success — each in uncooked efficiency metrics and in showcasing automated structure design — positions Liquid AI as one of many rising gamers to observe within the evolving AI mannequin panorama.
