Liquid AI, the Boston-based foundation model startup spun out of the Massachusetts Institute of Technology (MIT), is seeking to move the tech industry beyond its reliance on the Transformer architecture underpinning most popular large language models (LLMs) such as OpenAI’s GPT series and Google’s Gemini family.

Yesterday, the company announced “Hyena Edge,” a new convolution-based, multi-hybrid model designed for smartphones and other edge devices in advance of the International Conference on Learning Representations (ICLR) 2025.

The conference, one of the premier events for machine learning research, is taking place this year in Vienna, Austria.

New convolution-based model promises faster, more memory-efficient AI at the edge

Hyena Edge is engineered to outperform strong Transformer baselines on both computational efficiency and language model quality.

In real-world tests on a Samsung Galaxy S24 Ultra smartphone, the model delivered lower latency, smaller memory footprint, and better benchmark results compared to a parameter-matched Transformer++ model.

A new architecture for a new era of edge AI

Unlike most small models designed for mobile deployment — including SmolLM2, the Phi models, and Llama 3.2 1B — Hyena Edge steps away from traditional attention-heavy designs. Instead, it strategically replaces two-thirds of grouped-query attention (GQA) operators with gated convolutions from the Hyena-Y family.

The new architecture is the result of Liquid AI’s Synthesis of Tailored Architectures (STAR) framework, which uses evolutionary algorithms to automatically design model backbones and was announced back in December 2024.

STAR explores a wide range of operator compositions, rooted in the mathematical theory of linear input-varying systems, to optimize for multiple hardware-specific objectives like latency, memory usage, and quality.

Benchmarked directly on consumer hardware

To validate Hyena Edge’s real-world readiness, Liquid AI ran tests directly on the Samsung Galaxy S24 Ultra smartphone.

Results show that Hyena Edge achieved up to 30% faster prefill and decode latencies compared to its Transformer++ counterpart, with speed advantages increasing at longer sequence lengths.

Prefill latencies at short sequence lengths also outpaced the Transformer baseline — a critical performance metric for responsive on-device applications.

In terms of memory, Hyena Edge consistently used less RAM during inference across all tested sequence lengths, positioning it as a strong candidate for environments with tight resource constraints.

Outperforming Transformers on language benchmarks

Hyena Edge was trained on 100 billion tokens and evaluated across standard benchmarks for small language models, including Wikitext, Lambada, PiQA, HellaSwag, Winogrande, ARC-easy, and ARC-challenge.

On every benchmark, Hyena Edge either matched or exceeded the performance of the GQA-Transformer++ model, with noticeable improvements in perplexity scores on Wikitext and Lambada, and higher accuracy rates on PiQA, HellaSwag, and Winogrande.

These results suggest that the model’s efficiency gains do not come at the cost of predictive quality — a common tradeoff for many edge-optimized architectures.

Hyena Edge Evolution: A look at performance and operator trends

For those seeking a deeper dive into Hyena Edge’s development process, a recent video walkthrough provides a compelling visual summary of the model’s evolution.

The video highlights how key performance metrics — including prefill latency, decode latency, and memory consumption — improved over successive generations of architecture refinement.

It also offers a rare behind-the-scenes look at how the internal composition of Hyena Edge shifted during development. Viewers can see dynamic changes in the distribution of operator types, such as Self-Attention (SA) mechanisms, various Hyena variants, and SwiGLU layers.

These shifts offer insight into the architectural design principles that helped the model reach its current level of efficiency and accuracy.

By visualizing the trade-offs and operator dynamics over time, the video provides valuable context for understanding the architectural breakthroughs underlying Hyena Edge’s performance.

Open-source plans and a broader vision

Liquid AI said it plans to open-source a series of Liquid foundation models, including Hyena Edge, over the coming months. The company’s goal is to build capable and efficient general-purpose AI systems that can scale from cloud datacenters down to personal edge devices.

The debut of Hyena Edge also highlights the growing potential for alternative architectures to challenge Transformers in practical settings. With mobile devices increasingly expected to run sophisticated AI workloads natively, models like Hyena Edge could set a new baseline for what edge-optimized AI can achieve.

Hyena Edge’s success — both in raw performance metrics and in showcasing automated architecture design — positions Liquid AI as one of the emerging players to watch in the evolving AI model landscape.