NVIDIA releases Dory1.0-4B, a 4B looped language model with elastic inference compute

NVIDIA has released NVIDIA-Nemotron-Labs-Dory1.0-4B-MXFP8, a 4B-parameter looped language model with up to 256K context under the OpenMDW-1.1 license.

The model applies a shared recurrent block multiple times, letting users change loop count to trade inference compute for accuracy without retraining.

The card lists a last-loop KV caching mode that keeps the cache at one-loop size, and reports benchmark scores at 1, 2, and 4 loops.