NVIDIA opens Kumo foundation models for tabular business data and RFD4-Proteina, a protein design model with the Baker Lab it says is 13x faster than RFdiffusion3

Why it matters

【Key points】

NVIDIA announced new open models at GTC Berlin; two families are new.

1. Kumo: open-weight foundation models for tabular and relational business data, part of the new NVIDIA Agent Knowledge Foundation. Kumo Structured covers relational/tabular prediction, forecasting and anomaly detection; Kumo Engine is an open library stack. Agents can forecast, score fraud and recommend without a separate ML pipeline per task. Coinbase, Databricks, DoorDash, Revolut and Tubi are early users; on Hugging Face and build.nvidia.com.

2. BioNeMo RFD4-Proteina: an open protein design model from NVIDIA and the Baker Lab (UW Institute for Protein Design) that generates sequences and 3D structures; NVIDIA says it leads on 10+ design tasks and is ~13x faster than RFdiffusion3 on 200-residue proteins. On GitHub, Hugging Face and build.nvidia.com.

3. Also: Cosmos post-training tools and Cosmos Data Factory; Alpamayo 2 Edge due later this month; new Nemotron Retriever models; Ising on AWS PCS. Nemotron 3.5 Bolt E4B 4GB still due Oct 23.

【Verifiable】

NVIDIA blog: https://blogs.nvidia.com/blog/open-models-europe-industry/

【Our read】

1. Background: NVIDIA ships open models almost weekly (Nemotron 3.5 Super VL, Cosmos3-H-Dreams, Flora on this site). Kumo and RFD4-Proteina stand out: one brings open weights to the tabular data enterprises actually run on, the other merges the RFdiffusion line with NVIDIA's Proteina.

2. Comparison: tabular foundation models were mostly academic (TabPFN) or closed services; DeepMind's AlphaProteo remains closed, so RFD4-Proteina could become the most accessible high-end protein design tool.

3. Impact: free models, paid compute. Kumo inside Databricks and Snowflake could pull classic ML workloads onto GPU inference. Tabular and science models are thin spots in Chinese open-source, worth watching for responses.

4. Caveats: the 13x and "leading" claims are self-reported; the blog gives no Kumo sizes, licence or benchmarks—check the model cards.