NVIDIA used the Hub blog to announce Kumo Tabular, a small open model that reads a labeled table and predicts the next rows. EverMind's Raven paper, dated Sep 27, landed on Daily Papers this morning. Fastino's 340M GLiNER2.5-Decide has been on the Hub since last week for anyone who wants labels without a generated paragraph.
On the Hub
NVIDIA Kumo Tabular
NVIDIA announced Kumo Tabular on the Hub blog on Sep 29. The model is a Transformer for tabular classification and regression, with sizes from 28M to 215M, under the OpenMDW-1.1 license. The Hub cards have been up since Sep 1.
Give it a table of labeled rows and it predicts labels for new rows in a single forward pass. There is no training loop, no tuning, and no feature engineering. NVIDIA says it was pretrained only on artificial data, with column, row, and in-context attention in the TabICL and TabPFN style. The blog claims first place on TabArena, BeyondArena, TALENT, and ScoringBench.12
Raven
EverMind's Raven paper hit Hugging Face Daily Papers today. The paper itself is dated Sep 27 (arXiv 2609.33439). It describes an open multi-agent system that builds and evolves modular harnesses around models instead of asking one model to do the whole job.
A Host Agent breaks a goal into subtasks, matches each one to a specialized agent, and keeps the dependencies in order. A host archive and EverOS hold experience across runs, and Skill Forge turns that experience into reusable procedures. This is a paper on the Daily Papers list, not a new weights drop.3
fastino/GLiNER2.5-Decide
Fastino's GLiNER2.5-Decide is a 340M Apache-2.0 English classifier. You pass any label set at call time (intent, routing, sentiment, priority, policy) and it answers in one forward pass with no generated tokens. The card went up on Sep 23.
Fastino claims 60.2 percent exact-match on its in-house fast-decisions suite, covering 17 domains with 300 held-out examples each. It is built for operational routing, and it does not reason or explain.4
