Artificial intelligence liquid revolutions in LLMS to work on edge devices like smartphones with the new “Hyena Edge” model

Artificial intelligence liquid, which is a foundation company based in Boston, outside the Massachusetts Institute of Technology (MIT), is seeking to transfer the technology industry beyond its dependence on transformer structure that supports the most common large language models (LLMS) like LLMS like Openai’s GPT Series and Gemini from Google family.

Yesterday, the company announced “Edge of hyena“A new model depends on wrapping, multi -spell International Conference on Learning Representatives (ICLR) 2025.

The conference will be held, one of the leading events of automated learning research, this year in Vienna, Austria.

A new model based on wrapping is faster and more efficient in memory on the edge

Hyena Edge is designed to outperform strong transformer lines on both mathematical efficiency and the quality of the language model.

In the real world tests on the Samsung Galaxy S24 smartphone, the model delivered less transmission time, smaller memory stigma, and better results compared to the ++ transformer model that matches the teachers.

The new architecture for a new era of the brink of artificial intelligence

Unlike most small models designed to spread the mobile phone-including SMOLLM2, Phi Models, and Llama 3.2 1B- Steps of hyena are steps away from traditional designs to pay attention. Instead, strategists replace two -thirds of the attention of the gathering (GQA) with walled orders of the Al -Dabaa family.

The new architecture is the result of the synthesis of liquid from the artificial intelligence of the specially designed structure (Star), which uses the evolutionary algorithms to automatically design the model spine and It was announced in December 2024.

Star explores a wide range of operator’s works, rooted in the sporty theory of variable systems of linear inputs, to improve multiple goals for devices such as cumin and use memory and quality.

Directly on consumer devices

To check Hyena Edge ready in the real world, AI conducted the liquid directly on the Samsung Galaxy S24 Ultra smartphone.

The results showed that Hyena Edge has previously achieved up to 30 % in advance and deciphering the code compared to his ++ counterpart, with increased advantages of speed with longer sequence lengths.

Prosecution operations in short sequence lengths also exceeded the transformer basic line-an important performance scale for the applications of the insignificant device.

Regarding memory, Hyena Edge has been constantly used less quantities of RAM during inferring all tested sequence lengths, and put them as a strong candidate for narrow resource restrictions.

Overcoming transformers on language standards

Hyena Edge has been trained on 100 billion icons and was evaluated through the standard standards of small language models, including Wikitext, Lambada, Piqa, Hellaswag, Winogrande, Arc-SYY and Arc-Callenge.

In each standard, the edge hyena corresponds to either or exceeds the performance of the GQA-TRARANSFORMER ++, with noticeable improvements in the score on Wikitext and Lambada, and higher accuracy rates on PIQA, hellaswag and Winogrande.

These results indicate that the gain of the efficiency of the model does not come at the expense of predictive quality-a common comparison of many improved structures.

Hyena Edge Evolution: A look at performance and operator trends

For those who seek the deepest diving in the process of developing Hyena Edge, which is a modern Walk with the video It provides a convincing visual summary of the development of the model.

The video highlights how to improve the main performance measures – including prior transition time, decipher, and memory consumption – over successive generations of improving architecture.

It also provides a rare view behind the scenes on how the inner formation of the edge of the hyena turns during development. Viewers can see dynamic changes in the distribution of the types of operators, such as self -interest mechanisms (SA), various hyena variables, and Swiglu layers.

These transformations provide an insight into the principles of architectural design that helped the model reach the current level of efficiency and accuracy.

By imagining the bodies and the dynamics of the operator over time, the video provides a valuable context for understanding the architectural breakthroughs behind Hyena Edge’s performance.

Open source plans and a wider vision

Amnesty International said it is planning to open a series of liquid basic models, including Hyena Edge, in the coming months. The company’s goal is to build capable and effective artificial intelligence systems that can expand from cloud data centers to personal edge devices.

Hyena Edge’s first appearance highlights the increasing potential of alternative structures for the challenge of transformers in practical environments. Since mobile devices are increasingly expecting complex work burdens of artificial intelligence, models like Hyena Edge can put a new essential line of what improved artificial intelligence can achieve.

The success of Hyena Edge – in raw performance standards and the presentation of automatic architecture design – AI as one of the emerging players watching in the advanced model scene.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.