AI Today: Kolibri and NVIDIA Kumo Relational

State of knowledge as of 2026-10-26. Source material gathered between 2026-10-22 and 2026-10-26.

Today we examine two significant solutions in the intelligent systems sector. The first is Kolibri, a specialized mixture-of-experts model optimized for sovereign domains and the German language. The second novelty is NVIDIA Kumo Relational, a solution designed to work directly on relational databases, eliminating the need for manual data preparation for analytical and predictive processes. Both tools offer a new approach to performance in their respective categories.

Highlights Today (TL;DR)

• Kolibri: An open 78.1B parameter model optimized for data sovereignty and the German language.
• NVIDIA Kumo Relational: A tool for working on databases with a scale of 500B rows without requiring model training.

Kolibri

On October 26, 2026, Kolibri was unveiled as an English-German mixture-of-experts model released under the Apache 2.0 license. It is an open-weights model featuring 78.1B total parameters, of which 3.46B parameters are activated per token (representing 4.4% of all parameters).

The model was built with sovereign and regulated domains in mind, such as public administration, industry, and aerospace. The creation process involved full control over architecture, data, and training infrastructure. Kolibri was trained on 24T tokens, with German making up over 20% of the data. To achieve this share, over 2T web-mined and synthetically generated tokens were utilized. The post-training process combines supervised fine-tuning with reinforcement learning on over 1.2M internally curated tasks covering reasoning, tool use, instruction following, coding, and information retrieval. According to published data, Kolibri sits on the Pareto frontier for quality and inference cost for English and German. The model's design builds upon an earlier iteration named Kolibri Origin.

For builders using AI: Technical teams working in regulated German-speaking sectors can test the model's performance in processes requiring data sovereignty. The source does not specify the availability of the Polish language in the trained model nor does it provide hardware requirements for running a model of this scale—these details must be verified before production deployment.

NVIDIA Kumo Relational

NVIDIA introduced the Kumo Relational foundation model on October 21, designed to work directly on databases without requiring a training process. This tool supports databases with over 500 billion rows and enables predictions in a single forward pass. The solution operates directly on the relational structure, accounting for tables, foreign keys, and timestamps, which eliminates the need to manually flatten data into a single table.

The model utilizes an in-context learning mechanism, gathering up to 10,000 historical examples to predict outcomes such as customer churn, demand, or fraud. In conducted tests, Kumo Relational outperformed traditional machine learning models and gradient-boosted trees across 41 tasks on 15 databases. On the SAP SALT benchmark, the solution achieved an 8% higher Mean Reciprocal Rank (MRR) than AutoGluon teams. The tool is available as an open-source library named structured-data-models or as an NVIDIA NIM service. Thanks to the Predictive Query Language (PQL), the model can be directly leveraged by AI agents to predict events and take actions.

For builders using AI: Data engineering teams and business analysts can use Kumo Relational to automate predictions without feature engineering. The source does not provide details on infrastructural requirements for local deployments, the costs of using the NVIDIA NIM service, or support for specific SQL dialects used in various database systems—these matters require verification before integration with existing production pipelines.

Frequently Asked Questions (FAQ)

What are the main use cases for the Kolibri model?

Kolibri was created for work in sovereign and regulated domains. Key application areas include public administration, industry, and aerospace, where full control over the model's architecture, data, and training infrastructure is required.

How does NVIDIA Kumo Relational process data?

The model operates directly on the relational structure, taking into account tables, foreign keys, and timestamps. By applying an in-context learning mechanism, it utilizes up to 10,000 historical examples, eliminating the need to manually flatten data into a single table.

Sources

• Kolibri: A Sovereign European Model on the Pareto Frontier — 2026-10-26
• NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required — 2026-10-21

--- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Version before human editing.