Icon_LLM
TildeOpen LLM

How TildeOpen LLM is built

A closer look at key moments, breakthroughs, and whatā€™s next.

Timeline_uz_tumsa_fona

Notes from the Lab šŸ“‹

Behind-the-scenes commentary, insights, and updates from our research team on TildeOpen development.

04.09.2025.

TildeOpen is now live on Hugging Face! After a year of research and development on Europeā€™s fastest supercomputers, our team has published TildeOpen. Itā€™s ready for you to explore, experiment with, and build on.

8.01.2026.

Recent benchmark results confirm TildeOpenā€™s strong performance in grammatical accuracy, reading comprehension, and efficiency across morphology-rich European languages. These results reinforce its suitability as a reliable foundational model for multilingual AI applications. View full benchmark results on Hugging Face and on TildeBench.

16.07.2025.

TildeOpen is nearing the end of its development! After seeing 2 trillion tokens, the foundational model has been trained, and weā€™re now moving on to fine-tuning and evaluation. Once ready, the fine-tuned models will be published on Hugging Face.

09.06.2025.

Weā€™re proud to be among the first companies to test JUPITER, Europeā€™s first exascale supercomputer! With 1.2 million GPU hours granted to us, weā€™ll adapt TildeOpen for real-world use ā€“ including multilingual enterprise search, context-aware assistants, and other secure AI tools.

27.05.2025.

Great news! Weā€™ve secured an additional 140,000 GPU hours on LUMI through EuroHPC JU. These resources will be used to instruction-tune the model as part of the FFplus-funded project, focusing on key multilingual tasks such as translation, summarisation, and question answering.

12.05.2025.
We are halfway there with the pretraining! Seeing one trillion tokens took longer than anticipated due to monkey patching bugfixes and waiting for GPU allocations.
05.05.2025
We’ve introduced a more efficient exampleā€‘packing strategy for supervised instruction tuning in EleutherAI’s GPTā€‘Neox. Early profiling shows roughly 90% packing efficiency, keeping LUMI’s GPUs almost as fully utilised as during pretraining. Another improvement is the multiturn instruction masking strategy, allowing the model to participate in lengthy multi-turn conversations.
15.04.2025
We’ve now completed roughly one-third of pretraining. Getting there meant hammering out a stack of quirks, bugs, and some truly artisan code in EleutherAI’s GPT-Neox ā€“ plus a couple of our own blunders. However, this required only a single, very early restart, so almost no GPU time was lost!
15.03.2025

We have finally started the long-awaited TildeOpen pretraining. Borrowing from Mark Twain: “Quitting smoking is the easiest thing in the world; I’ve done it thousands of times.” Let’s hope this run is not a false start and delivers the results we’ve been working towards for so long!

web_LLM

See how LLMs really perform

Created by our researchers, TildeBench is a public leaderboard tracking how various LLMs handle tasks like machine translation, in-context question-answering, and grammar-sensitive text generation ā€“ all in languages that are often overlooked. Itā€™ll be updated with new tasks and models over time.

Compare LLMs on TildeBench

Build AI that speaks your languageā€ÆĀ 

TildeOpen gives you the foundation to create secure and sovereign AI. Explore the model now
or talk to us about tailoring it to your needs.

Talk to us
Go to Hugging Face