Open Source at UBS

Open source plays an important role in UBS’s technology strategy. Across the firm, engineering teams contribute to open-source projects, participate in industry foundations, and collaborate with technology communities to help advance the tools and platforms used across financial services and the wider technology ecosystem.

UBS’s open-source activity spans AI, cloud-native technologies, developer platforms, data engineering, and sustainable software. UBS participates in and supports industry initiatives including FINOS, the Linux Foundation, the Cloud Native Computing Foundation, and the Green Software Foundation. These activities include foundation membership, participation in technical and governance forums, contributions to community-led projects, and the maintenance of publicly available open-source projects.

UBS also maintains and contributes to open-source projects such as sparkutils, Quality, Frameless, and Kogito-4-Spark, alongside contributions to wider community projects including Apache Spark, Scales XML, Terraform Azure providers, FINOS initiatives, and vLLM. These projects and contributions reflect how UBS both uses open-source technologies across the business and contributes improvements back to the communities that maintain them.

Related link

Expand all panels

Building an enterprise AI model portfolio: scaling self-hosted and open-source AI

As AI adoption accelerates, organizations need effective ways to understand, plan, and manage the resources required to run AI solutions responsibly, securely, and efficiently.

UBS is developing an approach that helps teams select and access the right AI models for different needs, from large general-purpose models to smaller, more specialized options. This supports a wide range of use cases while helping meet requirements around data residency, legal obligations, operational effectiveness, and UBS-specific model adaptation.

Expanding UBS’s model portfolio with self-hosted models

Many organizations access large language models (LLMs) through vendor-provided Application Programming Interfaces (APIs). These services can be effective for a broad range of AI use cases, but some workloads require tighter control over where data is processed and how it is handled. In certain cases, model inference may trigger in the cloud tenant owned by UBS but it can cross to the cloud operated by the model provider, or in other cases, it may remain within the cloud tenant of UBS but capacity may be optimized dynamically across different geographical locations, thus sending requests to different data centers across the globe. For highly sensitive use cases, including some in private banking, this can create data residency and governance considerations. To address these requirements, UBS is expanding its self-hosted AI capabilities. By running selected models within environments controlled by UBS, the firm can maintain stronger oversight of data flows, processing locations, and operational controls for sensitive information.

Easy and efficient self-hosting through vLLM

The initiative began with an assessment of whether open-weight small language models could be deployed effectively in UBS-operated Kubernetes clusters. To support this work, the Chief AI Office adopted vLLM, an open-source inference and serving engine for large language models. Originally developed at UC Berkeley, vLLM has since become a widely adopted, community-driven project supported by contributors across academia and industry.

Building on this early success, the capability was formalized as AI-as-a-Service (AIaaS). It has since expanded to support strategic initiatives with specialized requirements and has grown into an enterprise-wide platform trusted by hundreds of users, processing more than one billion tokens per day within the first month after launch.

To deliver an enterprise-grade service at UBS, AI-as-a-Service is also developing comprehensive internal benchmarking and exploring intelligent model routing. These capabilities help optimize inference by ensuring each request is directed to the most appropriate model for the task, while balancing performance, cost, and operational requirements. The following initiatives from the vLLM open-source community are being leveraged for this purpose:

GuideLLM: A platform for evaluating how language models perform under realistic workloads and configurations. It simulates end-to-end interactions with OpenAI-compatible and vLLM-native servers and generates workload patterns that reflect production usage.
vLLM Semantic Router: An intelligent request router that directs incoming queries to the best model for each task. Positioned between users and the vLLM Inference Engine, it can reduce computing cost and latency by routing simpler queries to faster models while reserving more complex tasks for heavyweight reasoning models.

Benefits beyond self-hosting

Consistent guardrails and controls

Business continuity

Vendor-provided models may be subject to non-negotiable decommissioning timelines, and access to specific models can also be suspended by governments due to national security concerns. In these circumstances, financial institutions such as UBS must follow business continuity policies and guidelines, mitigate vendor concentration risk, and maintain alternative solutions. By offering a portfolio of models with comparable capabilities and sizes through the same API, AI-as-a-Service enables AI use cases to develop fallback strategies if a model becomes unavailable, minimizing the impact of such scenarios.

Cost optimization

AI use cases are developed to deliver measurable business outcomes. Because AI systems are probabilistic, governance processes often require teams to demonstrate that model performance has been quantitatively evaluated and that the results meet an acceptable standard. Access to models with different characteristics, such as parameter counts, context-window sizes, reasoning capabilities, and cost profiles, enables AI engineers to optimize each use case across multiple dimensions. This allows teams not only to maximize performance, but also to make informed decisions about the trade-off between performance and cost.

Evaluation is key for model selection

The previous sections outlined several reasons why an AI use case may need to change its underlying model: stricter data residency requirements, the need for fallback options to support business continuity, or the pursuit of a better performance-to-cost balance. In each case, the performance of the AI use case must be measured, and the impact of any model change must be quantified.

To fully realize the benefits of a rich model portfolio, this capability must be complemented by a robust AI evaluation framework. Across the industry, both offline and online evaluations are becoming integral steps in the development and operation of agentic systems.

Offline evaluations: Offline evaluations take place before release in a controlled development environment. They assess how well an agentic or generative system performs a specific task using a prepared dataset of questions and expected answers. Comparing different models through comprehensive offline evaluations helps identify acceptable performance and cost trade-offs for each business case.
Online evaluations: Online evaluations take place after the agent is deployed and interacting with real users. They monitor live traffic to assess how the agent behaves in unpredictable, real-world scenarios. Significant and sustained performance degradation may require teams to re-evaluate the system and, where appropriate, replace or adjust the models behind it.

For this purpose, the Chief AI Office is assessing various open-source solutions such as MLflow or Langfuse.

Contributing back

Through these efforts, UBS continues to deepen its expertise in efficient AI infrastructure and responsible model deployment. Stability and robustness are essential requirements for enterprise-grade LLM hosting. Through extensive testing and simulated workloads, UBS engineers identified several bugs and improvement opportunities. Recently, AI engineers in the Chief AI Office submitted contributions to address these issues and strengthen GPU support for a range of open-weight LLMs used by UBS.

These contributions represent one example of UBS's broader engagement with the open-source ecosystem. By contributing improvements back to community projects, UBS helps strengthen the technologies it relies upon while supporting innovation across the industry.

Where appropriate, the Chief AI Office contributes improvements back to open-source communities, supporting broader innovation while enhancing the capabilities required for enterprise-grade AI at UBS.

Skip Links

Search Title

Open Source at UBS

Our approach

Learn more about technology at UBS

Technology careers

Investing in the future

Academic partnerships

Our firm

Media

Investors Relations

Career at UBS