NVIDIA open-sources Kumo Tabular: synthetic-data-trained foundation model tops 8 benchmarks
jure · x · 2026-11-19
NVIDIA released the technical report, training recipe, and weights for Kumo Tabular, a family of open tabular foundation models (arXiv/HF/GitHub).
- Three sizes: S (27M), M (61M), L (214M params), pretrained entirely on synthetic tables from the Kumo prior — zero training on new tables.
- Method: a table encoder compresses rows into embeddings; an in-context learning transformer predicts unlabeled rows; architectural innovations enable scaling to 150M cells.
- Results: Kumo Tabular-L ranks first on 8 benchmarks (TabArena, TALENT, BeyondArena, etc.); all sizes sit on accuracy-speed Pareto frontiers, with L predicting 36x faster than LimiX-2 on TabArena and 1.7x faster than TabPFN-3.5 on BeyondArena.
More from Infra
- NVIDIA driver 621.24 constantly writes to SSDs via NVDisplay.Container logging — giveen · 2026-11-19
- Intel Foundry bull case: TSMC's capacity crunch opens second-source opportunity — BenBajarin · 2026-11-19
- Firecrawl Local Goes Open Source: 37x Faster Rust Scraping, 11MB Idle Memory — gaganghotra_ · 2026-11-19
- Investor thesis: the future of compute is heterogeneous, many chips + routing — hargup13 · 2026-11-19
- Lambda deploys NVIDIA Photonics CPO switches: 10% lower latency, 15x higher reliability — TheZachMueller · 2026-11-19
- Intel clarifies Microsoft deal: EMIB packaging, Foundry custom silicon separate from Xeon — BenBajarin · 2026-11-19