Nemotron Labs Dory 1.0 (4B)

NVIDIA Nemotron Labs Dory 1.0 4B is an open-weight looped (latent-recurrent) normalized-GPT language model with interleaved sliding-window attention (SWA) and full-attention layers. A shared recurrent block can be applied multiple times at inference (recommended default: 2 loops) to trade compute for accuracy without swapping checkpoints. Text-in / text-out, English, up to 256K context (max_position_embeddings 262144). Reasoning model with tool calling. Pre-trained on 6T tokens (NVFP4 then BF16) plus 300B context-extension tokens; training data collected through December 2025. Released under OpenMDW License Agreement v1.1. Weights: BF16 (siblings MXFP8, NVFP4, GGUF).