KumoRFM-2
#311 in Frontier LLMsKumo Ai · v2 · since 14. April 2026 · 8× · last seen Jun 30, 2026
0
Momentum
KumoRFM-2 is a foundation model developed by Kumo AI for relational (tabular) enterprise data, not a classic text language model. It uses in-context learning and a hierarchical graph transformer architecture to generate predictions (e.g. churn, fraud, demand) directly on connected database tables without feature engineering or task-specific training. The model was announced on April 14, 2026, scales to over 500 billion rows, and reportedly outperforms both classical machine learning models and its predecessor on several benchmarks (RelBench, SAP SALT). Access is via a Python SDK (pip install kumoai) and a public inference API with natural-language querying (PQL).
Momentum trend
19.05.17.08.
Features
| Key Benchmark (%) | Outperforms strongest supervised ML model on Stanford RelBenchV1 by 5%; on SAP SALT benchmark reaches 0.89 mean reciprocal rank (fine-tuned) vs. 0.77 (AutoGluon) |
| Context Window (Tokens) | No text context window; scales to 500+ billion database rows, throughput up to 5 GB/sec and 20M lookups/sec |
| License | Proprietary model/SDK with API-key access (free key via kumorfm.ai, limit 1000 requests/day); accompanying 'Kumo Coding Agent' separately released under MIT license |
| Multimodality | Processes relational databases (tables as graph); nodes can carry numerical, categorical, timestamp, text, and vector embedding attributes |
| Platform | Python SDK (pip install kumoai, Python 3.10+), public inference API at kumorfm.ai, connects to SQL databases and cloud data warehouses (Snowflake, Databricks, Spark) |
| Release Date | April 14, 2026 |