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



AI Summary

Overview

  • Liquid AI, a Boston startup from MIT, launched Hyena Edge on April 25, 2025.
  • Designed to run AI more efficiently on mobile devices.

Key Features

  • Hyena Edge uses a convolution-based multi-hybrid model, reducing reliance on traditional transformer architecture.
  • Replaces 2/3 of attention operations with gated convolutions, improving speed and memory efficiency.
  • Proven to operate faster and with less memory on a Samsung Galaxy S24 Ultra compared to traditional transformer models.

Performance Metrics

  • Hyena Edge showed up to 30% faster performance with lower memory usage across various token lengths.
  • Maintained or improved accuracy metrics compared to transformer baselines across several standard benchmarks:
    • Perplexity on wiki text dropped from 17.3 to 16.2.
    • Lambda accuracy improved from 10.8 to 9.4.

Architectural Evolution

  • Developed using an evolutionary algorithm named STAR, which optimized various model configurations over multiple generations.
  • Final model consists of 32 layers, width of 48, and 64 attention heads, with significant automation in its design process.

Future Prospects

  • Liquid AI plans to open-source Hyena Edge, inviting further development and adaptations for various devices.
  • This signals a shift towards hybrid models that balance performance with efficiency, especially for edge devices, moving us into a post-transformer era.