Kumo Ai · od 20. Mai 2025 (KumoRFM); KumoRFM-2 am 14. April 2026 angekündigt · 7× · naposledy 30. 6. 2026
0
Momentum
KumoRFM is a Relational Foundation Model (RFM) by Kumo AI for predictions on structured, relational business data (e.g., from data warehouses). It converts database tables internally into temporal, heterogeneous graphs and uses in-context learning with a Relational Graph Transformer to make predictions (classification, regression, link prediction) without task-specific training. Queries are made via a SQL-like Predictive Query Language (PQL); access is via Python SDK, API, and MCP server. In April 2026, the successor KumoRFM-2 was introduced, scaling to up to 500 billion rows.
Vývoj momenta
19.05.17.08.
Vlastnosti
| Key Benchmark (%) | Outperforms feature-engineering and deep-learning baselines on RelBench (30 tasks) by 2–8% on average; additional 10–30% improvement after fine-tuning |
| Context Window (Tokens) | No classical token context window; scales to 500B+ rows of relational data via graph engine with memory-mapped I/O |
| Multimodality | Processes numerical values, categories, timestamps, text, and vector embeddings per table row/node |
| Platform | Python SDK (pip install kumoai, Python 3.10+), API, MCP server, integrates with Snowflake/Databricks/AWS and data warehouses |
| Release Date | May 20, 2025 (KumoRFM-1); April 14, 2026 (KumoRFM-2) |