NVIDIA Kumo Tabular ships open weights. The Elo does not get a second vote.

NVIDIA published NVIDIA Kumo Tabular on September 29, 2026, as an open tabular foundation model on Hugging Face. It comes in three sizes, from 28 million to 215 million parameters, under OpenMDW-1.1, with code in NVIDIA/structured-data-models. It predicts labels from labeled rows in a single forward pass, and it was pretrained only on artificial tables.

TL;DR - On NVIDIA Kumo Tabular, NVIDIA says the model is first on TabArena at an Elo of 1950, and 17 times faster than LimiX-2 on one RTX 6000 Pro. - BeyondArena, in NVIDIA's account, is an Elo of 1418 and an Improvability of 7.78%. A single pass covers up to 10 classes. - Training stages go to 60,000 rows and 100 columns. Tech AI Wire rewrites the post. It is not a second measurement.

What actually happened

Tech AI Wire's same-day piece repeats the headline numbers. It does not independently verify the Elo.

The product claim is easier to picture than a leaderboard. You have labeled rows. The model predicts labels from those rows in one forward pass.

That is in-context tabular prediction as NVIDIA describes it, not a story about a long training job on your warehouse every time you ask a question. The pass covers up to 10 classes. If your label set is larger than that, this sentence is a wall, and the launch post is the thing that built it.

The pretraining line should be in the first slide, not the appendix. NVIDIA says the model was pretrained only on artificial tables. Artificial tables are not your customers, your factory logs, or your billing export.

Open weights do not change the diet the weights saw. They change whether you can download the result. Those are different facts, and only one of them is a data claim.

Sizes run from 28 million to 215 million parameters. The license named on the release is OpenMDW-1.1. The code location named with it is NVIDIA/structured-data-models.

On NVIDIA Kumo Tabular, this piece does not expand that license into a list of rights the pack does not spell out. If legal wants the grant language, they need the license text, not a blog recap of a blog recap.

The ranking claim is NVIDIA's. It says the model is first on TabArena at an Elo of 1950, and 17 times faster than LimiX-2 on one RTX 6000 Pro. BeyondArena is reported at an Elo of 1418, with Improvability at 7.78%.

Training stages go to 60,000 rows and 100 columns. Every one of those figures is coming from the company that shipped the model. A higher Elo, in this article, means NVIDIA's ranking claim. It does not mean a referee you can name from a second lab showed up in the pack.

Tech AI Wire, also dated September 29, repeats the 1950 Elo, the 17 times speed line, the 28 million to 215 million size range, and the date. The page is a rewrite of the NVIDIA post. A rewrite can be accurate as copying and still add zero evidence.

The second page does not independently verify the Elo. If a meeting treats "two outlets" as confirmation, this pair does not qualify.

What to do as a reader

Put a source column next to every number, and watch how fast the column collapses to one cell. Elo 1950, the 17 times comparison on one RTX 6000 Pro, BeyondArena 1418, Improvability 7.78%, the 10-class cap, the 60,000-row and 100-column training stages, the artificial-table pretraining, the three sizes, the license name, and the code repo are NVIDIA's post. Tech AI Wire is useful as a mirror.

It is not a second experiment. Say that out loud when someone waves the mirror.

Do not promote "first" past TabArena as NVIDIA scored it. First at 1950 is a claim about that board. BeyondArena at 1418 is a different board's number in the same post, and this article will not tell you how to convert one Elo into the other.

The pack does not include the conversion, the opponent list, or the match count. Inventing a plain-English "how much better" from 1950 would be a new number. Do not.

Treat the artificial pretraining as the pilot constraint. A fair next step, if this model is even relevant to your tables, is to hold out real rows the model never saw and see whether a single forward pass, within the 10-class limit, does anything you can check. That is a reader method, not a promise that it will work.

The launch does not report your data. It reports artificial tables, then a leaderboard.

On NVIDIA Kumo Tabular, hardware belongs in the same caution. The 17 times figure is against LimiX-2 on one RTX 6000 Pro, as NVIDIA states it. It is not a blanket speed claim for every GPU, every row count, or every class count under the cap.

The training-stage limits, 60,000 rows and 100 columns, are about how the model was staged, and they are not a substitute for a measured pass on your file. If your table is wider or taller than that description, you are outside the sentence they published.

Open weights still matter for a narrow reason. You can look at the Hugging Face release and the code in NVIDIA/structured-data-models instead of trusting a screenshot of an Elo. Looking is not the same thing as having verified 1950.

It is the step that keeps the rewrite from becoming the source. If the weights and the post disagree, the post's leaderboard does not win by being repeated.

What a second source would have to be

A second source would name its own runs, its own hardware, and whether it rebuilt the TabArena setup. Tech AI Wire does not do that in the account used here. It repeats 1950, repeats 17 times, repeats the parameter range, and repeats September 29.

Agreement between a post and its rewrite is the weakest form of agreement there is. It rules out some kinds of misquoting. It does not rule out a vendor-only score.

Improvability at 7.78% and the BeyondArena Elo are easy to drop because they are less catchy than "first." They are still part of the same NVIDIA note, and they keep the story from being a single medal. The 10-class limit does the same job on the product side.

A foundation model for tables that stops at 10 classes in one pass is a specific tool. Calling it a general replacement for every tabular pipeline is a sentence the source did not write.

None of this is a reason to buy hardware, switch a stack, or treat a leaderboard as a forecast. The practical residue is a checklist. Confirm the date, September 29, 2026.

On NVIDIA Kumo Tabular, confirm the sizes and the license name on the Hugging Face post. Confirm that any Elo you repeat is NVIDIA's, unverified by the second page. Then decide whether artificial-table pretraining and a 10-class pass even match the problem in front of you.

The pages are NVIDIA's Hugging Face post and Tech AI Wire's rewrite. Cite the first for the claims. Cite the second only to say it does not independently verify the Elo.

Readers who want sourced recaps that are already on the site can read DogeOS public testnet and PUMP token buyback burn as separate live posts.