ModelsPublished All news from that day
Nvidia listed Dory 1.0 4B, a looped model with 256K context and up to 61.8 on SWE-bench Verified.
Nvidia's model card for Dory 1.0 4B describes a looped model whose loop count, from one to four, trades inference compute for accuracy.
Check our sources · 6 facts from 1 sourceNvidia published Nemotron Labs Dory 1.0 4B on Hugging Face with a general-availability date of 16 October 2026. The model card describes a looped (latent-recurrent) language model with 4B parameters that can repeat a shared block of layers at inference time, and a context length of up to 256K tokens.
The card names the OpenMDW License Agreement, version 1.1, and says the model is ready for commercial or non-commercial use. Its stated design goal is letting users trade inference compute for accuracy by changing the loop count, from one loop up to four.
The card lists SWE-bench Verified scores of 50.3, 59.5 and 61.8. These are Nvidia's figures for 1, 2 and 4 loops, which Nvidia says it measured under a consistent harness (NeMo Gym / Nemo Evaluator SDK).
Underlined sentences link to the facts we checked. Dotted words open a short explainer.
Fact 4 of 6 · from Hugging Face / NVIDIA
This model is ready for commercial or non-commercial use. (OpenMDW License Agreement, version 1.1)
Fact 5 of 6 · from Hugging Face / NVIDIA
By changing how many times the shared block of layers is applied, you can trade inference compute for accuracy to suit each task, with no retraining and no model swap.
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against this source.
1 NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 model card
Open the source archived copy-
The Dory 1.0 4B model card describes a looped (latent-recurrent) language model with 4B parameters that can repeat a shared block of layers at inference time.
citeis a looped (latent-recurrent) large language model (LLM) trained by NVIDIA ... By repeatedly applying a shared block of layers ... Number of model parameters: 4B
-
The model card's version list shows GA (10/16/2026).
citeGA (10/16/2026)
-
The model card gives a context length of up to 256K tokens.
citeUp to 256K tokens
-
The model card cites the OpenMDW License Agreement, version 1.1, and says the model is ready for commercial or non-commercial use.
citeThis model is ready for commercial or non-commercial use. (OpenMDW License Agreement, version 1.1)
-
The model card says the design goal is to let users trade inference compute for accuracy by changing the loop count, from one loop up to four.
citeBy changing how many times the shared block of layers is applied, you can trade inference compute for accuracy to suit each task, with no retraining and no model swap.
-
The model card lists SWE-bench Verified scores of 50.3, 59.5 and 61.8 for Dory1.0-4B at 1, 2 and 4 loops, which NVIDIA says it measured under a consistent harness (NeMo Gym / Nemo Evaluator SDK).
Toned down to what the source says
citeSWE-bench Verified | 50.3 | 59.5 | 61.8 under Dory1.0-4B 1 loop, 2 loops, 4 loops. Accuracy numbers measured by NVIDIA under a consistent harness (NeMo Gym / Nemo Evaluator SDK).
Every sentence above is checked against this source. Each fact has its own address you can cite.
Source 1 · Hugging Face / NVIDIA
NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 model card
The Dory 1.0 4B model card describes a looped (latent-recurrent) language model with 4B parameters that can repeat a shared block of layers at inference time.
is a looped (latent-recurrent) large language model (LLM) trained by NVIDIA ... By repeatedly applying a shared block of layers ... Number of model parameters: 4B
Show all 6 factsShow fewer facts
The model card cites the OpenMDW License Agreement, version 1.1, and says the model is ready for commercial or non-commercial use.
This model is ready for commercial or non-commercial use. (OpenMDW License Agreement, version 1.1)
The model card says the design goal is to let users trade inference compute for accuracy by changing the loop count, from one loop up to four.
By changing how many times the shared block of layers is applied, you can trade inference compute for accuracy to suit each task, with no retraining and no model swap.
The model card lists SWE-bench Verified scores of 50.3, 59.5 and 61.8 for Dory1.0-4B at 1, 2 and 4 loops, which NVIDIA says it measured under a consistent harness (NeMo Gym / Nemo Evaluator SDK).
Toned down to what the source saysSWE-bench Verified | 50.3 | 59.5 | 61.8 under Dory1.0-4B 1 loop, 2 loops, 4 loops. Accuracy numbers measured by NVIDIA under a consistent harness (NeMo Gym / Nemo Evaluator SDK).
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep