Liquid AI is revolutionizing LLMs to work on edge devices like smartphones with new ‘Hyena Edge’ model

Misachusetts Institute of Technology (MIT) Massachusetts Technology Institute (MIT) The beginning of the liquid AI, the most popular large language models (LLS) want to move outside the transformer architecture Openai’s GPT The series and Google’s Twins The family.

The company announced yesterday “Hyena edge“It is designed for a new convolution based, very hybrid model, previous smartphones and other edge devices International Conference on Learning Representations (ICLR) 2025.

The conference, one of the first measures for machine learning research, takes place in Vienna, Austria this year.

The new convolutionary model, faster than the edge, more memory promises an EU

Hyena Edge is designed to advance both calculation efficiency and powerful transformer baselines in the quality of the language model.

A Samsung Galaxy S24 is a better benchmark results compared to real world tests, models, small memory traces, smaller memory traces, smaller memory tracks and better benchmark + model.

A new architecture for a new period of AI

Smollm2, PHI Models and Llama 3.2 1B – Hyena Edge, unlike most small models, unlike the smallest models that are far from traditional attention-heavy designs. Instead, the attention of the inquiries of the Hyena-y’s secret congresses (GQA) is a third of the operators strategically replaced.

The new architecture is the result of a special architectural (star) framework, which uses the synthesis of liquid EU’s synthesis of liquid EU to automatically design the model waist, and Declared back in December 2024.

To optimize multiple hardware for specific purposes, such as stars, delays, memory usage and quality, the linear entrance is an extensive operator composition tuned into the mathematical theory of various systems.

Directly viewed to the consumer’s apparatus

To confirm the real world training of Hyena EDGE, liquid AI direct tests on the Samsung Galaxy S24 Ultra Smartphone.

The results show that the Hyena shows that the delays are delayed in advance of up to 30% faster and compared to the transformer ++ counterpart, and the delays are delayed and delays open.

Delay delays in short sequence lengths, transformed base – a critical performance metric for applications applied on the device.

In memory of memory, Hyena Edge, consistently, used the less Aries during the inference along the length of all test sequence, place it as a strong candidate for the environment with a hard welding.

Transformers in language criteria

Hyena Edge has developed 100 billion signs and rated in standard criteria for small language models, including Wikitext, Lambada, Piqa, Hellaswag, ARC-Easy and Arc-Challenge.

In each benchmark, in the edge of Hyena, or exceeded the performance of the GMA-Transformer ++ model, which adapted to a portable degree and higher accuracy rates in Wikitext and Lambada, Hellaswag and Winogrande.

These results show the lack of prediction quality of the effectiveness of the model – the presence of a general trade for many edge optimized architecture.

Hyena Edge Evolution: A look at performance and operator trends

For those who want a deeper dive to the development process of Hyena Edge Video Walkthrough Provides a compulsory visual summary of the evolution of the model.

https://www.youtube.com/watch?v=n5al1jlupca

The video is improved on consecutive generations of architectural elegance, including basic performance measurements – pre-delay, coded and memory consumption.

In addition, a rare scenic is designed behind a rare scenes that the internal content of Hyena changes during the development. Viewers can see the dynamic changes in the distribution of operator types such as self-focus (SA) mechanisms, various Hiyena options and swig.

These turns provide information on the principles of architectural design that helps the model reach the current efficiency and accuracy.

By visualizing trade-offs and operator dynamics over time, the video offers a valuable context to understand the architectural progress based on Hyena Edge’s main performance.

Open source plans and a wider vision

The liquid AI said that in the coming months, including Hyena Edge, including Hyena Edge, including Hyena Edge. The purpose of the company is to create a skillful and effective general purpose AI systems that can be scale up to individual edge devices from Cloud Datocenters.

The debut of the Hyena’s edge also protests transformers in practical parameters for alternative architecture. The models such as mobile devices, which are still developed by developed AI workload, can build a new foundation for the most popular models such as Hyena Edge.

Hyena Edge’s success – as one of the players that form both raw performance dimensions, as well as in automated architectural designs and liquid AI positions, developing AI model view.