Liquid AI revolutionize LLMS to operate on end devices such as smartphones with a new Hyena Edge model
Join our daily and weekly newsletters for the latest updates and exclusive content of a leading AI coverage industry. Learn more
The Liquid AI, Boston-based Foundation Model Launches by the Massachusetts Institute of Technology (MIT), seeks to move the technology industry beyond its reading of transformer architecture that underlies the most popular models in large languages ​​(LLM) OPENAI GPT series Google twins Family.
The company announced yesterday “Hyena Edge“New, based on a revolution, a multi -hybrid model designed for smartphones and other end devices before International Conference on Training of Training (ICLR) 2025.
The conference, one of the best events to study machine learning, is being held this year in Vienna, Austria.
The new Convolution based model
Hyena Edge is designed to outperform strong base transformers both computing and the quality of the linguistic model.
In the real world tests of the Samsung Galaxy S24 Ultra smartphone, the model delivers lower latency, a smaller memory footprint and better comparison results than a ++ parameter.
New Architecture for a new era of Edge AI
Unlike most small models designed for mobile implementation-influential Smollm2, Phi models and Llama 3.2 1b-Hyena Edge move away from traditional heavy attention designs. Instead, he strategically replaces two-thirds of the operators of the Group’s attention (GQA) with closed blankets from the Hyena-Y family.
The new architecture is the result of the Liquid AI Synthesis Framework of Adapted Architectural (Star), which uses evolutionary algorithms for automatic design of base models and was announced as early as December 2024.
Star explores a wide range of operator compositions, rooted in the mathematical theory of linear input systems to optimize for multiple hardware -specific targets such as latency, memory use and quality.
Directly directly on consumer hardware
To validate the readiness of Hyena Edge in the real world, tests for liquid test tests directly on the Samsung Galaxy S24 Ultra smartphone.
The results show that Hyena Edge has achieved up to 30% faster delays and delays decoding compared to his transformer ++ colleague, with the benefits of speed increasing at longer sequence lengths.
Preliminary delays at short lengths of the sequence also outpaced the baseline of the transformer-critical indicator of the effectiveness of the responsive applications of the device.
In terms of memory, Hyena Edge consistently uses less RAM during a conclusion for all tested lengths of sequence, positioning it as a strong candidate for an environment with strict resource restrictions.
Superior transformers of language indicators
Hyena Edge was trained in 100 billion tokens and estimated through standard small languages ​​models, including Wikitext, Lambada, Piqa, Hellaswag, Winograde, Arc-Easy and Arc-Challenge.
Each Hyena Edge indicator either matches or exceeds the operation of the GQA-Transformer ++, with noticeable improvements in Wikitext and Lambada’s bewilderment and higher PIQA, Hellaswag and Winograph accuracy.
These results suggest that increasing the efficiency of the model does not come with the cost of the estimated quality-a general compromise of very optimized architectures.
Hyena Edge Evolution: A look at the performance and trends in the operator
For those looking for a deeper diving in the process of developing Hyena Edge, recently Video step Provides an overwhelming visual summary of the evolution of the model.
The video emphasizes how key performance performance – including delay in preferences, latency of decoding and memory consumption – has improved through the successive generations of architectural advancements.
It also offers a rare backstage view of how Hyena Edge’s internal composition is shifting during development. Viewers can see dynamic changes in the distribution of operators, such as self -recognition mechanisms (SA), different hyena variants and Swiglu layers.
These shifts offer an idea of ​​the architectural principles of design that have helped the model reach its current level of efficiency and accuracy.
By visualizing the compromises and dynamics of the operator over time, the video provides a valuable context for understanding the architectural breakthroughs that underlie the presentation of Hyena Edge.
Open code plans and wider vision
The Liquid AI said it plans to open a series of models on liquid foundations, including Hyena Edge, in the coming months. The aim of the company is to build capable and effective general -purpose AI systems that can scale from cloud data centers to personal devices.
Hyena Edge’s debut also emphasizes the growing potential for alternative architectures to cause transformers in practical settings. With mobile devices that are increasingly expected to perform complex AI loads, models such as Hyena Edge can set a new baseline for what AI-optimized with EDGE AI can achieve.
The success of Hyena Edge – both in raw performance indicators and in demonstrating automated architecture design – positions Liquid AI as one of the emerging players to watch in the AI ​​Model developing landscape.
English