AI news · October 18, 2026
NVIDIA's Dory makes loop count a runtime choice
Dory1.0-4B lets builders change from one to four inference loops without a new checkpoint; NVIDIA reports 61.8 on four-loop SWE-bench Verified.
- parameters
- 4B
- context tokens
- 256K
- four-loop SWE-bench Verified
- 61.8
NVIDIA released Nemotron Labs Dory1.0-4B on October 16 as a small English model for general reasoning. Its normalized-GPT design repeats a shared block at inference, so one checkpoint can spend more or less compute on a request. The model card lists a 256K-token context, commercial and non-commercial use, and BF16, MXFP8, NVFP4, and Q5_K_M formats.
NVIDIA's table reports 50.3, 59.5, and 61.8 on SWE-bench Verified at one, two, and four loops. The card recommends two loops as a starting point and four for harder agent tasks, while noting that current vLLM support is under review. That makes this a runtime tuning story: the useful result is the best loop count for one workload, not the largest score in the table.
What you can do with it
Serve the same coding task with one, two, and four loops. Record accepted patches, tool errors, output tokens, GPU memory, first-token time, and total time. Start with two loops, then keep a higher count only when the extra accuracy is worth the added compute.
Our take
Dory gives a small model a useful control: spend more compute only when the task needs it. The model card's scores are vendor-run, and support is still moving in some serving tools. A three-arm local comparison can show if the control improves accepted work rather than only benchmark results.
Links Try it
Source: NVIDIA / Hugging Face — Made With Models writes the brief; the reporting is theirs.