Research
Google Research unveils Muir, a linear attention architecture for long context
Muir reports streaming 10M tokens in a 25 MB state without perplexity drift and 90.0% zero-shot passkey accuracy at 1M tokens, claiming gains over Gated DeltaNet-2 and a 2.7x training speedup over Splash Attention at 32K. The results come from a Reddit post linking an arXiv paper.