As the tech landscape continually evolves, a significant transition is underway led by a Boston-based startup, Liquid AI. Spinning off from the prestigious Massachusetts Institute of Technology (MIT), this startup is daring to challenge the prevailing Transformer architecture that has dominated the development of large language models (LLMs). At the heart of their innovation lies the recently announced Hyena Edge, a groundbreaking convolution-based model tailored for smartphones and edge devices, signaling a decisive move towards a more efficient and capable future for artificial intelligence.
Hyena Edge is unveiled ahead of the International Conference on Learning Representations (ICLR) 2025, a prestigious gathering for machine learning minds held in Vienna, Austria. In a world where tech giants like OpenAI and Google have heavily relied on a Transformer framework, Liquid AI’s approach stands as a beacon of change. What is even more striking is the model’s promise to outperform its Transformer counterparts in both computational efficiency and language understanding, positioning Hyena Edge as a formidable contender in the AI domain.
A New Breed of AI Architecture
The impressive capabilities of Hyena Edge extend far beyond simple marketing claims. Tested on the Samsung Galaxy S24 Ultra, this model showcases an astonishing decrease in latency and memory usage, disrupting industry norms. In a landscape littered with small models like SmolLM2 and Llama 3.2 1B, Hyena Edge dares to break free from the confines of conventional attention-heavy methods. Instead, it introduces a revolutionary mix by replacing two-thirds of traditional grouped-query attention (GQA) operators with advanced gated convolutions derived from the Hyena-Y family.
Liquid AI’s proprietary Synthesis of Tailored Architectures (STAR) framework is the backbone of this innovation. By employing evolutionary algorithms, STAR empowers researchers to automatically curate model architectures targeted at addressing key performance metrics, including latency and memory efficiency. This method diverges from conventional processes, offering a dynamic and data-driven approach to constructing high-performance AI solutions.
Real-World Performance: A Game-Changer
Liquid AI’s commitment to practicality is evident through their rigorous validation of Hyena Edge’s real-world applicability. The performance tests conducted on the Samsung Galaxy S24 Ultra reveal that Hyena Edge can accelerate prefill and decode latencies by up to 30% compared to the Transformer++ model. This advancement is crucial for creating responsive AI applications where every millisecond counts.
Moreover, the memory consumption during inference consistently showcases a lower RAM requirement across all tested sequence lengths. Such efficiency positions Hyena Edge as an ideal candidate for devices operating under stringent resource constraints. The model was trained on a staggering 100 billion tokens, and its performance on standard benchmarks, including Wikitext and PiQA, confirms that it can hold its ground or even excel beyond its Transformer counterparts.
Benchmarking Excellence
Perhaps one of the most compelling aspects of Hyena Edge is its ability to deliver superior predictive capabilities without sacrificing efficiency. It outshines various benchmarks, exhibiting improved perplexity scores and accuracy across a range of datasets. Whether it’s Wikitext or Lambada, Hyena Edge consistently surpasses expectations, offering insights into a future where mobile devices are equipped to handle complex AI workloads seamlessly.
Liquid AI’s transparency about its development process further adds credibility to its claims. A recent video walkthrough offers an in-depth look at how Hyena Edge evolved through generations. By showcasing the shifts in operator types, such as Self-Attention mechanisms and different Hyena variants, the video illustrates the underlying engineering principles driving the model’s performance. This nuanced perspective on architectural development not only informs but also inspires future innovations in the AI space.
A Commitment to Open Source and Innovation
In a move that reflects its dedication to advancing the AI landscape, Liquid AI plans to open-source its foundation models, including Hyena Edge. This strategic decision sets a precedent for a collaborative approach to AI development, facilitating a broader sharing of knowledge and technological prowess. By democratizing access to its models, Liquid AI aims to cultivate a robust ecosystem where capable and efficient AI systems flourish from cloud datacenters to personal edge devices.
Hyena Edge’s debut not only marks a critical milestone for Liquid AI but also signals a growing recognition of the potential for alternative architectures in the AI domain. As devices continue to integrate sophisticated AI capabilities, models like Hyena Edge pave the way for a future where edge-optimized AI becomes the new standard. The time for innovation is now, and with platforms like Liquid AI at the forefront, the horizon is bright for both technology developers and end-users alike.
Spanish 
English