← Back to Futurice and UpCloud build a self-hosted sovereign RAG agent on Qwen3 and vLLM, deployed entirely in Finland

Source proof for Futurice and UpCloud build a self-hosted sovereign RAG agent on Qwen3 and vLLM, deployed entirely in Finland

Source-bound proof

Verified source excerpts for every supported field

Each colour maps a published value to the exact source passage used to support it. Only bounded excerpts are public; administrators can inspect the complete captured source.

12 fields supported

Title

Futurice and UpCloud build a self-hosted sovereign RAG agent on Qwen3 and vLLM, deployed entirely in Finland

quote · high
…t is a practical starting point for identifying capability gaps and priorities. Together with UpCloud, we moved beyond a simple assessment form and built a self-hosted RAG-based AI agent deployed entirely in Finland, demonstrating sovereign AI capabilities. This blog shares the technical details of the architecture, GPU infrastructure,…
…ssible while keeping enough capacity for parallel processing of multiple users. For the models, we stopped at the Qwen3 family, which has both inference and embedding models: Qwen3-30B-A3B-Instruct-2507-FP8 - mid-size inference model with 30 billion para…

Description

Futurice, together with UpCloud, built a self-hosted RAG-based AI agent for a digital-sovereignty self-assessment, deployed entirely in Finland on managed Kubernetes with an Nvidia L40S GPU, n8n as the agent runtime, and Qwen3-30B and Qwen3-Embedding models served with vLLM. The post reports load testing with 60 parallel users at an average response time below 20 seconds and a cost comparison against GPT-5-mini on Azure.

quote · high
…e sovereign AI automation platform concept we described previously in our blog. This architecture relies on a managed Kubernetes service on UpCloud and n8n as an AI agent runtime and co-working platform. UpCloud provides servers with Nvidia L40S GPUs, which are capable of running qu…
…iss). During the test, we were collecting metrics from Prometheus for analysis. 60 user amount was found to be optimal, because it allows to maintain stable average response time below 20 seconds in both scenarios. Details are presented below on the test result visualisation. As presented, the…
…0 output tokens per request, with an average response time of about 18 seconds. Let’s compare our Qwen3-30B model to a non-thinking GPT-5-mini on Azure with consumption-based pricing. We take this model because it was launched about the same time as Qwen3 and has…

Company

Futurice

quote · high
…be Family companies Columbia Road Fram Partners Meltlake Qlarify Recordly Thriv We are part of the Futurice family © 2026 Futurice. All rights reserved Futurice Privacy Policy Impressum

Country

Finland

quote · high
…priorities. Together with UpCloud, we moved beyond a simple assessment form and built a self-hosted RAG-based AI agent deployed entirely in Finland, demonstrating sovereign AI capabilities. This blog shares the technical detail…

Industry

Professional Services

classification · medium
…setup for GPU infrastructure for backup and scaling, along with latency tests. Get in touch to discuss sovereign AI solutions - Pass the sovereignty assessment - Author Anton Andreev Tech Principal anton.a…

Problem

Digital sovereignty and sovereign AI are crucial for EU organisations needing secure, compliant AI infrastructure, and a self-assessment is a practical starting point; Futurice wanted to go beyond a simple assessment form and demonstrate sovereign AI capabilities.

quote · high
…reign AI RAG solution with self-hosted LLMs and UpCloud 3 Jun 2026 | Technology Digital sovereignty and sovereign AI are crucial for EU organizations needing secure, compliant AI infrastructure. A self-assessment is a practical starting point for identifying capability gaps…

Solution

Futurice and UpCloud built a self-hosted RAG AI agent deployed entirely in Finland on a managed Kubernetes service on UpCloud with an Nvidia L40S GPU, using n8n as the agent runtime, a Next.js frontend, PostgreSQL with PGVector as knowledge store, and Qwen3-30B and Qwen3-Embedding models served with vLLM.

quote · high
…orking. Check out the referenced blog post for more details – in this solution, the only addition was an extra simple Next.js-based frontend which connects to n8n via webhooks. Retrieval-Augmented Generation (RAG) fundamentals After the foundations are sel…
…etes service on UpCloud and n8n as an AI agent runtime and co-working platform. UpCloud provides servers with Nvidia L40S GPUs, which are capable of running quite powerful AI solutions. For our use, UpCloud has all the managed services we need for running a reliabl…

Business value

In load testing with 60 parallel users the system kept average response time below 20 seconds, and the authors' cost comparison against GPT-5-mini on Azure found self-hosting becomes economically meaningful at about 14,700 requests a day. This is a self-demonstration system, not a client deployment.

quote · high
…Each request would cost 0.22/1M 9000/2 + 0.03/1M 9000/2 + 1.74/1M*400=0.001821. Considering these numbers, a self-hosted solution becomes economically meaningful already at 14700 requests a day, which is similar to 9 users polling the system without stopping 8 hours a day.…
…trained on the same topic-related materials and our internal company knowledge. With this setup, we aimed to demonstrate that modern systems can be deployed close to the business location. The system is available to be seen in action at sovereignty.futurice.com. Build…

AI capabilities

Retrieval-Augmented Generation, Generative AI, Conversational AI

classification · high
…priorities. Together with UpCloud, we moved beyond a simple assessment form and built a self-hosted RAG-based AI agent deployed entirely in Finland, demonstrating sovereign AI capabilities. This blo…

Technology

n8n, Qwen3, vLLM, Kubernetes, UpCloud

classification · high
…e sovereign AI automation platform concept we described previously in our blog. This architecture relies on a managed Kubernetes service on UpCloud and n8n as an AI agent runtime and co-working platform. UpCloud provides servers with Nvidia L40S GPUs, which are capable of running qu…

Headline outcome

derived · high
…iss). During the test, we were collecting metrics from Prometheus for analysis. 60 user amount was found to be optimal, because it allows to maintain stable average response time below 20 seconds in both scenarios. Details are presented below on the test result visualisation. As presented, the…

Use case type

Conversational assistant

classification · medium
…, the customer can receive a detailed report with risks and recommendations, or discuss the results with the AI chatbot trained on the same topic-related materials and our internal company knowledge. With this setup, we aimed to demonstrate th…
Capture details
Captured
16 Nov 2026, 06:03 UTC
Extractor
fetch-strip@1
Snapshot hash
3b5bda51adc019e56e0a913fd4b9d73911b7a473b413eb9897a6f8affbffc18e