Is NVIDIA Kumo state of the art?Pending benchmarks.Announced Oct 21, 2026; no independent results yet.

The details so far

Lab

NVIDIA

Announced

Oct 21, 2026

API

Not yet

Open weights

Published

Parameters

Kumo Tabular 28M / 62M / 215M (Small / Medium / Large); Kumo Forecast ~340M frozen MOMENT-1-large encoder plus a trained head; Kumo Anomaly 8M; Kumo Relational not stated

Models with weights

nvidia/Kumo-Tabular, nvidia/Kumo-Relational, nvidia/Kumo-Forecast and nvidia/Kumo-Anomaly on Hugging Face

Licenses

Kumo Tabular and Kumo Relational weights: OpenMDW 1.1. Kumo Forecast and Kumo Anomaly: Apache 2.0

Kumo Tabular

In-context tabular Transformer (column, row and in-context attention), pretrained only on synthetic tables from structural causal models; up to 10 classes per forward pass. Weights and blog published Sep 29, 2026

Kumo Relational

Based on KumoRFM-2; fine-tunes TabICLv2 weights (BSD-3-Clause) and adds relational message passing; also served as a Kumo Relational NIM

Kumo Forecast / Anomaly

Run with the Kumo-TS library. Forecast pairs a MOMENT encoder with retrieval of historical patterns (DARR); Anomaly is diffusion-based with adaptive thresholds

Libraries

structured-data-models (Apache 2.0, GitHub) for Tabular and Relational; Kumo Engine, an open library stack for building and deploying structured-data models

Snowflake

NVIDIA says Snowflake customers can use Kumo Engine and Kumo Relational in Snowflake Notebooks; NVIDIA's docs include a Snowflake Notebooks guide

Kumo AI

Kumo Relational is based on KumoRFM from Kumo AI. Fortune and The Information reported in June 2026 that NVIDIA had acquired Kumo AI; neither company formally announced itunsure · press

"Kumo Time Series"

Named in one NVIDIA blog post; no model card or weights under that name. NVIDIA's docs list Forecast and Anomaly as the time-series models (run via Kumo-TS)unsure

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: Kumo Tabular ranks first on TabArena (Elo 1950), BeyondArena, TALENT and ScoringBench, and runs 17x faster than LimiX-2 on a single RTX 6000 Pro.

Lab claim: Cognite, Coinbase, Databricks, DoorDash, Revolut and Tubi are among the first to work with Kumo models; Databricks pairs Kumo Anomaly with Databricks Genie Ontology.

Sources