organisation · org/nvidia
NVIDIA
Also called NVIDIA Corporation, Nvidia, Nemotron
The company that sells the hardware gives away the parts of a model that are
usually the secret. The card for nvidia/nemotron-3-ultra-550b-a55b —
550B total and 55B active, on a Mamba-2, mixture-of-experts and attention hybrid with multi-token predictionsource, accessed 2026-08-28 — says that
"major portions of the pre-training corpus are released"
as a public dataset collection, that major portions of the fine-tuning corpus
are released too, and that the end-to-end training recipe is published in
NVIDIA's developer repository. All ten of NVIDIA's rows in the OpenRouter
catalog, as observed on 31 August 2026, carried a Hugging Face id, and no
vendor with more rows than that published weights for all of them: the next
fully-open listings belonged to meta-llama and MoonshotAI, tied at eight
rows each.
The licence became a standard one as the models stopped being someone else's.
NVIDIA's April 2025 reasoning release was a Llama derivative under
the Nvidia Open Model Licensesource, accessed 2026-08-28; the Nemotron 3 family announced
15 December 2025 is NVIDIA's own hybrid architecture under
the OpenMDW License Agreement, version 1.1source, accessed 2026-08-28, an off-the-shelf agreement rather than a
house one. The exception is the interesting row.
nvidia/nemotron-3.5-content-safety, the guardrail model that screens inputs
and responses for other systems, is
a fine-tune of Google's Gemma-3-4B-it, governed by OpenMDW plus the Gemma Terms of Use and Gemma Prohibited Use Policysource, accessed 2026-08-28. NVIDIA's safety model inherits a
competitor's acceptable-use policy, and anyone deploying it to enforce
policy is bound by Google's.
Then there are the rows that argue with each other. Five of the eighteen free
listings in the catalog, as observed on 31 August 2026, were NVIDIA's — more
than any other vendor — and
two of them advertise a larger window than the paid row of the same model:
nvidia/nemotron-3.5-lightning lists
262144tokensopenrouter-models, last checked 2026-09-14 while
nvidia/nemotron-3.5-lightning:free lists
1000000tokensopenrouter-models, last checked 2026-09-14, a figure
the public page repeats. Nemotron 3 Ultra does the same;
nvidia/nemotron-3-super-120b-a12b runs the other way, its free row at
262144tokensopenrouter-models, last checked 2026-09-14 against
the paid row's
262144tokensopenrouter-models, last checked 2026-09-14. And
nvidia/nemotron-3-ultra-550b-a55b:batch heads
higher than the row it batches on both input and output —
not publishedlast known value, as of an unrecorded date — the source no longer lists this rowopenrouter-models, last checked 2026-09-14 against
$0.60per million tokensopenrouter-models, last checked 2026-09-14 in, and
not publishedlast known value, as of an unrecorded date — the source no longer lists this rowopenrouter-models, last checked 2026-09-14 against
$2.40per million tokensopenrouter-models, last checked 2026-09-14 out, where the
convention is a discount. Neither figure is necessarily NVIDIA's: each is the
top listed provider's rate for its row, and two rows are not obliged to be
headed by the same provider, so the inversion sits between two listings
rather than being a surcharge anyone levied. It does come with a longer
window:
not publishedlast known value, as of an unrecorded date — the source no longer lists this rowopenrouter-models, last checked 2026-09-14 against
262144tokensopenrouter-models, last checked 2026-09-14.
Facts
- founded
- 5 April 1993, by Jensen Huang, Chris Malachowsky and Curtis Priemsource, accessed 2026-08-28
- headquarters
- Santa Clara, Californiasource, accessed 2026-08-28
- manufacturing model
- fabless and contracting: NVIDIA designs its products and contracts all phases of manufacturing — wafer fabrication, assembly, testing and packaging — to outside supplierssource, accessed 2026-09-12
- wafer foundries
- its semiconductor wafers are produced by foundries including Taiwan Semiconductor Manufacturing Company Limited (TSMC) and Samsung Electronics Co., Ltd. (Samsung)source, accessed 2026-09-12
- model license
- the OpenMDW License Agreement, version 1.1source, accessed 2026-08-28
- previous model license
- the Nvidia Open Model Licensesource, accessed 2026-08-28
- flagship parameters
- 550B total and 55B active, on a Mamba-2, mixture-of-experts and attention hybrid with multi-token predictionsource, accessed 2026-08-28
- released artifacts
- weights, major portions of the pre-training and fine-tuning corpora, and the end-to-end training recipesource, accessed 2026-08-28
- guardrail base model
- a fine-tune of Google's Gemma-3-4B-it, governed by OpenMDW plus the Gemma Terms of Use and Gemma Prohibited Use Policysource, accessed 2026-08-28