Introducing “Domyn Small”: An Open European Reasoning LLM Built for AI Ownership

With support for over 50 languages and optimization for English and the main European languages — French, German, Italian, Spanish, Portuguese — Domyn Small was trained on the high-performance computing capabilities of CINECA's Leonardo Supercomputer. The release builds on Domyn and CINECA's ongoing LLM development collaboration, which began in 2024, and marks a new milestone in Domyn's mission to deliver fully sovereign, fully owned AI for regulated industries. This partnership also sits within the Italian AI Factory (IT4LIA) initiative, which is part of a broader European network of AI Factories aimed at transforming the national and European AI landscape through advanced infrastructure and a comprehensive portfolio of services available to research groups, public administrations, companies and startups.

“CINECA within the EuroHPC framework contributes to the mission of placing sovereign, open, high-performance computing at the service of Italian and European industry, research, and public administration”, said CINECA President Francesco Ubertini. “Domyn Small was trained on a public government funded European infrastructure hosted here in Italy.  That is not a detail — it is the foundation on which trustworthy, AI Act-compliant, European AI is built. CINECA's enabling role in this project is a demonstration that HPC public infrastructure and industrial AI ambition are not in competition but they can work in synergy for a European AI strategy which is both competitive and open”. ‍

A model for resource-constrained environments

Following the recent launch of the 260B-parameter-model Domyn Large, Domyn Small extends the family with a compact reasoning model designed for a different kind of deployment. Domyn Small is purpose-built for resource-constrained environments where compute, latency, and cost must be tightly managed — from on-premise deployments and regulated infrastructure to laptops and edge devices. The model offers an expandable context window and a togglable chain-of-thought reasoning mode (thinking on/off) that enables efficient multi-step reasoning across mathematics, coding, as well as tool calling for agentic applications. Domyn Small is comparable to peer models of the same size, exhibiting a positive trade-off between token efficiency (fewer tokens required per task) and accuracy on reasoning benchmarks. The result is lower cost and latency in production without sacrificing capability.

‍

AI built for sovereignty and ownership

Building on Domyn’s commitment to sovereign AI, this model will be released with open weights. This will allow companies to not only access, deploy, and run the model, but also modify its weights and training recipe to fine-tune it to a specific domain or use it as a base for new models.

"Domyn Small shows that Europe can lead in advanced AI by building reasoning models that are open, sovereign, and ready for production", said Uljan Sharka, CEO of Domyn. "Through our long-standing work with CINECA, we are delivering performance, efficiency, and full governance to the enterprises that depend on them most — those operating in regulated industries".

By combining open weights, sovereign infrastructure, and a model purpose-built for constrained, regulated, and multilingual environments, this release further enables enterprises and research institutions across Europe and beyond to build AI systems they can own, govern, and trust.