NVIDIA releases Nemotron Labs Dory1.0-4B with adjustable inference loops
The 4B English reasoning model has a 256K-token context and repeats a shared normalized-GPT block to vary compute with one checkpoint. Developer results cited in reporting put SWE-bench Verified scores at 50.3, 59.5, and 61.8 for one, two, and four iterations. The weights are available in several precisions for commercial and noncommercial use.
Visit original source