Dory1.0-4B Uses Latent-Recurrent Architecture for Elastic Inference Compute
November 1, 2026
Dory1.0-4B is a 4B-parameter nGPT model with a shared recurrent block that allows for variable inference compute via a single checkpoint. Pre-trained on 6 trillion tokens with a 256K context window, it utilizes a difficulty-aware loop curriculum during RLVR post-training.
HOW THIS AFFECTS YOU
●
builderYou can deploy a single model checkpoint that scales compute dynamically based on inference budget.
●
researcherThe interleaved sliding-window and full-attention architecture offers a new way to implement elastic depth.