- Who it helps
- Developers evaluating small local models and researchers comparing inference strategies.
- Availability
- The October 16 release includes downloadable weights under OpenMDW 1.1. The card lists BF16, MXFP8, NVFP4 and GGUF variants. Its vLLM support was still a pull request under review when checked; a routine package install may not support it yet.
What the loop changes
The same recurrent layers run more than once. NVIDIA recommends two loops as a starting point. Four loops did not improve every reported benchmark, so a larger loop count needs its own accuracy and latency check.
Loop count and context length also affect memory. The card distinguishes per-loop key/value caches from a last-loop cache. Those cached intermediate results help a model continue a conversation, but can occupy substantial memory beyond the model weights.
A useful first comparison
Choose ten short tasks with answers you can check: five exact data extractions and five reasoning problems. Keep their wording, output limits and quantization fixed. Compare one and two loops, recording correct answers, wall-clock time and peak memory. Include one deliberately unanswerable item to check whether the model invents missing data.
Save the precise runtime revision and model file. An eight-gigabyte estimate for raw BF16 weights does not describe the full memory needed to serve the model. A comparison is useful only when both configurations finish within your actual hardware limits.
What you can do
- Read the compatibility and cache sections of the model card before choosing a runtime or downloading a quantized file. Open reference ↗
Where the evidence stops
The benchmark table describes NVIDIA’s configuration. No local benchmark, quantization comparison or inference run was performed for this article.
Sources read. The full release card and community pointer were read on October 17 UTC. Available artifacts and documented constraints were checked; the model was not run.
Sources & further reading
Hugging Face · official
Nemotron Labs Dory 1.0 4B model card ↗
Full model card read in the browser after the web reader failed. Runtime support, loops and licensing were checked.
NVIDIA · page checked 2026-10-17 · Published 2026-10-16
Reddit · discussion
Community notice of NVIDIA Dory 1.0 ↗
Public post supplied the model-card link. It is a release pointer, with no independent benchmark established.
coder543 · page checked 2026-10-17
Coverage note: Primary and linked sources are limited to those listed. Independent execution, original X coverage and comparable cross-platform measurements were not established in this run.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
Dory 1.0 is NVIDIA’s four-billion-parameter text model with a repeated internal computation loop. It offers a way to trade more work per generated token for different reasoning performance.