New AI HYENA Destroys Old AI Models and Breaks Speed and Memory Records



AI Summary

Hyena Edge Overview

  • Liquid AI, a Boston startup from MIT, launched Hyena Edge on April 25th, 2025.
  • The model aims to improve AI performance on mobile devices, targeting faster and lighter operations.

Key Innovations

  • Hyena Edge utilizes a convolution-based multi-hybrid model instead of the traditional Transformer architecture.
  • It replaces approximately 2/3 of attention operations with gated convolutions.
  • Tested on a Samsung Galaxy S24 Ultra, it shows up to 30% faster prefill latency and lower memory usage than conventional models.

Performance Metrics

  • Compared to baseline transformer models, Hyena Edge exhibits significant improvements in perplexity scores across multiple benchmarks:
    • Wiki text: from 17.3 to 16.2
    • Lambada: from 10.8 to 9.4
    • Pyap: from 71.1 to 72.3
    • Hella swag: from 49.3 to 52.8
    • Wino Grande: from 51.4 to 54.8
    • Ark Easy: from 63.2 to 64.4
    • Arc Challenge: from 53.34 to 55.2
  • Both models tied on Pi QA variant with a score of 31.7.

Architecture Development

  • The design was derived using STAR (Synthesis of Tailored Architectures Framework), an evolutionary algorithm that evolved models over 24 generations.
  • Hyena Y operator balanced expressive power and efficiency, leading to the final Hyena Edge architecture.
  • The model features 32 layers deep, a width of 48, attention head size of 64, and reduced GQA blocks.

Future Implications

  • Liquid AI plans to open-source Hyena Edge, promoting accessibility and further development in the AI community.
  • Edge devices like smartphones will benefit from faster, more efficient AI, reducing reliance on cloud processing and enhancing privacy.
  • The evolution of AI architecture indicates a shift towards hybrid models and automated designs for performance improvements.