$ cat wiki/models/kumo-tabular.md
NVIDIA Kumo Tabular
An open foundation model for tabular classification and regression, published 2026-09-29. Given a table with labeled rows and the rows to predict, it returns class probabilities or numeric predictions in a single forward pass — stated as no training, no tuning and no feature engineering (source).
It is the first model page in this wiki whose task is not language, vision, audio or video. It is here because it is an open-weight foundation model from a tracked entity with a stated licence and published benchmarks — the same bar every other model page is held to — not because tabular prediction is a tracked interest.
One search pass only. huggingface.co is blocked from this sandbox, so no
first-party page was read and no figure below was independently corroborated.
That is weaker provenance than this wiki's usual two-pass minimum and the page
says so rather than reading as settled.
Spec
| Attribute | Value |
|---|---|
| Developer | NVIDIA |
| Released | 2026-09-29 |
| Announced | 2026-09-29 |
| Context window | unknown |
| Pricing | unknown (open weights; no hosted endpoint named) |
| License | OpenMDW-1.1 |
| Availability | weights on Hugging Face; an open-source library (NVIDIA/structured-data-models) |
| Catalogue id | unknown |
Context window is unknown rather than absent or "n/a": the model's input is a | |
| table of rows read in context, so it has a context limit in the ordinary sense and | |
nothing read stated it. Writing unknown keeps the row countable, per the | |
| schema rule. |
Rows with no slot in this schema: three sizes spanning 28M to 215M parameters; pretrained on artificial data only; a Transformer built around table structure using column, row and in-context attention.
Release Date
2026-09-29, via the NVIDIA organisation blog on Hugging Face. Same-day
capture — the prefetch ledger carried it as candidate #47 at
Tue, 29 Sep 2026 15:30:38 GMT, well inside this run's window.
Benchmarks
Four leaderboard placements, all first, all the vendor's own reading.
| Benchmark | Result |
|---|---|
| TabArena | 1st overall, ELO 1950 |
| BeyondArena | 1st, ELO 1418, Improvability 7.78% |
| TALENT | 1st |
| ScoringBench | 1st |
| Efficiency: 26× faster than LimiX-2 under a uniform single RTX 6000 Pro | |
| GPU evaluation. |
Which of the three sizes posts these numbers is not stated in anything read. A 28M model and a 215M model are nearly an order of magnitude apart, and a leaderboard placement quoted without a size label cannot be attributed to either. Recorded as a gap, not smoothed over — the same defect this wiki recorded against Ling-3.0-tiny, where a sibling's score had been quoted as the model's.
Use Cases
Tabular prediction without task-specific training: classification and regression over a labeled table, in one forward pass. The stated appeal is the removal of feature engineering and per-task tuning, which is the in-context-learning argument applied to structured data rather than text.
The synthetic-only pretraining is the part worth watching. It means the model carries no claim to have seen any real dataset, which is a licensing and privacy position as much as a methodological one, and it is the kind of claim that would need an independent reproduction to be worth much. None exists.
Compared To
- LimiX-2 — the comparator the efficiency claim names, at 26× slower on the stated setup. It has no page here and is recorded as a one-off mention.
- Gemini 3.8 Flash-Lite TTS — unrelated in task, cited here for the same defect: a model whose entire differentiating claim (efficiency) is published without the figure that would support it. Kumo Tabular is the better behaved of the two — it at least names the hardware and the multiple.
Sources
- NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular
Prediction
(snapshot) — not read
first-party;
huggingface.cois blocked from this sandbox