UCSD and Together AI Research Introduces Parcae: A Stable Architecture for Looped Language Models That Achieves the Quality of a Transformer Twice the Size
UCSD and Together AI just built a looped language model that matches a transformer twice its size. Here's why efficiency is the new arms race.
Why it matters
As inference costs dominate AI economics and models move to edge devices, a new architecture challenges the scaling orthodoxy—matching larger models with dramatically fewer parameters. This signals a fundamental shift from 'bigger is better' to 'smarter is cheaper.'
The key facts
5 to knowParcae architecture achieves transformer-equivalent quality at 50% parameter count
Research addresses inference-dominant compute allocation (vs. training-first paradigm)
Looped recurrent design enables edge deployment and reduced VRAM requirements
Co-authored by UCSD and Together AI research teams
Published April 2026 — recent architecture innovation in LLM design space
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: The dominant recipe for building better language models has not changed much since the Chinchilla era: spend more FLOPs, add more parameters, train on more tokens. But as inference deployments consume an ever-growing share of compute and model deployments push toward the edge, researchers are…