LumiOpen
LumiOpen publishes language models from a collaboration involving AMD Silo AI, the University of Turku's TurkuNLP group and HPLT. Its work develops Finnish and other European language capabilities. LumiOpen is presented here as a research collaboration, not as a separately verified corporation.
Company & leadership
- Organization
- Joint language-model research and publishing collaboration.
- Collaboration announced
- 28 August 2023, by the University of Turku and SiloGen.
- Contributors
- AMD Silo AI, TurkuNLP and the High Performance Language Technologies project.
- Training infrastructure
- LUMI supercomputer, with compute support from CSC in Finland.
- Industry participant
- Silo AI became part of AMD; AMD announced completion of the acquisition on 12 August 2024.
How the collaboration began
The University of Turku and SiloGen announced a joint model-development effort in August 2023. It connected academic language research with Silo AI's applied research team. The announcement identified Filip Ginter and Sampo Pyysalo as university project leaders and described HPLT as a source of European-language data.
Poro's training was complete by the February 2024 announcement. The project combined Finnish with English and code rather than relying on a Finnish-only corpus. Viking later extended this approach to several Nordic languages. The models are published through LumiOpen, while LUMI is the supercomputer used to train them.
Contributors and organizational changes
AMD announced its acquisition of Silo AI in August 2024, in a transaction valued at approximately $665 million. The team joined AMD's Artificial Intelligence Group. The announcement identifies Poro and Viking among Silo AI's model-development work. This acquisition concerns the industry participant; it does not make the University of Turku or HPLT AMD subsidiaries.
The 2026 Poro 2 Long release continues the AMD Silo AI and TurkuNLP partnership. It links the work to OpenEuroLLM's European language-model programme. The public project sources do not establish a separate LumiOpen incorporation date, headquarters or chief executive.
Research releases and scope
This profile covers Poro, Viking, long-context and mathematical reasoning releases, plus a masked diffusion experiment. Coverage remains incomplete. The official catalog also publishes the Llama-HPLT-edu family of language-specific educational-content classifiers. Those trained models still need individual records and review of their parent models, training data and release differences.
Practical guidance. This page contains the shared Poro and Viking model records. TurkuNLP's separate profile covers its own earlier Finnish releases without copying these records. Training-progress branches, GGUF packages and equivalent precision formats remain grouped with their underlying model.