Apple's ParaRNN Breakthrough: A 665x Training Speedup, the First 7B Classical RNNs, and the End of the Transformer Monoculture - April 26, 2026

Apple's ParaRNN Breakthrough: A 665x Training Speedup, the First 7B Classical RNNs, and the End of the Transformer Monoculture Apple just landed an oral at ICLR 2026 with ParaRNN, a framework that finally lets classical recurrent networks train in parallel and reaches transformer-competitive language modeling at 7 billion parameters. Rick Spair and Laura unpack the technical insight, the inference economics, the on-device strategy, and the legitimate skepticisms — and why this might be the m...