The language model from JKU spin-off Ablatic AI, announced in August, has been available to businesses since 15 September 2026. Talos-1 is designed for tasks where artificial intelligence does not answer questions, but carries out work: reading contracts, accessing interfaces, filling in forms, and completing a process across multiple systems. The model was trained on European high-performance computers as part of the EuroHPC programme and AI Factory Austria’s AI:AT initiative, and is operated in data centres in Austria and across the EU.
Tested in more than 100 companies
Between the announcement and the market launch, there was a two-and-a-half-month trial period involving more than 100 companies. The tasks ranged from marking homework on a learning platform, to searching through the internal documentation of a telecommunications company, to posting receipts in the accounts department.
“It’s a pleasure to work with the team at Ablatic. It’s important to develop high-performance AI for businesses in Europe, and that’s exactly what Ablatic delivers.”
- Manfred Strasser, Gateway Labs GmbH
An architecture that doesn’t become unmanageable as text length increases
At the technical heart of Talos-1 lies a model architecture designed in-house by Ablatic AI. In conventional language models, computational cost increases in proportion to the square of the text length: a document twice as long costs four times as much. Talos-1 breaks this trend. The model requires up to 14 times less computational power and moves up to 79 times less data through memory when generating a response than a conventional model of the same size. The longer the document, the greater the advantage: with 256,000 tokens – equivalent to around 500 pages – the advantage is eightfold; with one million tokens, it is 14-fold. These figures are calculated based on the architecture.
For a team the size of Ablatic AI, this is the difference between what is possible and what is impossible. Training on long sequences – which, for large providers, keeps data centres busy on a scale of billions – was thus feasible within the computing time allocated via EuroHPC. In operation, the same curve determines how many customers a server can serve simultaneously and what a request costs. At market launch, Talos-1 processes 200,000 tokens at a time; an increase to 512,000 is planned for this year, with one million to follow. The model itself is designed for one million; expansion is limited solely by the available server capacity.
“It was clear to us from the outset that we first had to reduce development and operating costs if we were to be competitive.”
– Julian Mauerkirchner, CTO and co-founder, Ablatic AI
On a par with agent-based tasks
In public benchmarks that test language models as autonomous agents, Talos-1 matches the performance of models from the US and China that are many times larger. On Toolathlon, a test for tasks in real-world software such as email, calendars and spreadsheets, Talos-1 scores 74.4 per cent; Anthropic’s Claude Opus 5 scores 80.6 per cent, whilst OpenAI’s GPT-5.6 Sol scores 74.9 per cent. When it comes to precisely following written instructions (IFBench), Talos-1 leads with 82.2 per cent, ahead of GPT-5.6 Sol, DeepSeek V4 Pro and GLM-5.3. When parsing real-world documents containing tables and formulas (OmniDocBench), it achieves 90.8 per cent, two points behind the top score. All test results, comparison models and sources are published at ablatic.ai/models.
“Ablatic AI demonstrates what a strong founding team can achieve by drawing on a high-performing innovation ecosystem centred around the university. Florian Zimmermann and Julian Mauerkirchner think ambitiously, implement ideas quickly and make targeted use of networks and contacts to further their development. I am convinced that this team has great potential to provide relevant impetus for European AI from Linz.”
- Elisabeth Ulbrich, CEO, LIT Open Innovation Centre (JKU)
European infrastructure, European law, one euro per million tokens
Talos-1 is operated exclusively by Ablatic AI and provided via an OpenAI-compatible interface, meaning that existing applications can be migrated to it without modification. The model processes text, images, video and files and responds in text; three levels of reasoning depth can be selected for each query. Customer data is not used for training; Ablatic AI acts as a data processor in accordance with Article 28 of the GDPR. The price is one euro per million input tokens processed, four euros per million output tokens generated and ten cents per million input tokens cached.
“Ablatic AI demonstrates just how much a team can achieve when ambition meets the right infrastructure: thanks to access to European supercomputer resources, made possible by AI Factory Austria, a model has been developed in a short space of time that is competitive on an international level and is taking on real work within companies. These are exactly the kind of results Europe needs right now.”
– Karl Kugler, Co-Lead, AI Factory Austria AI:AT
Access
Companies can contact Ablatic AI via ablatic.ai and will receive a response within one working day. The process begins with a pilot project based on the company’s own use case. Self-registration for the interface is planned as soon as server capacities have been expanded.
“I’ve been in IT since 1982. I’ve seen technologies come and go, and almost always there were people with a vision behind them – people whom perhaps someone still remembers today. In Julian Mauerkirchner and Florian Zimmermann, I have found two young Austrian visionaries who are not afraid of the future because they want to help shape it. Something great could emerge from ablatic, and that deserves to be supported.”
– Christian Haase, formerly of Pierer Innovation
Further information: www.ablatic.ai
About Ablatic AI
Ablatic AI is a spin-off founded in 2025 at the LIT Open Innovation Centre at Johannes Kepler University Linz. With Talos-1, the company is developing a European language model for use in businesses and regulated sectors. Ablatic AI was founded by Florian Zimmermann and Julian Mauerkirchner. Development, operation and data processing take place on European infrastructure and are governed by European law.
Regarding the announcement of 14 August 2026:
JKU spin-off Ablatic AI develops European - AI Factory Austria AI:ATAbout the AI Factory Austria AI:AT
AI Factory Austria AI:AT is Austria’s national AI Factory within the framework of the EuroHPC Joint Undertaking. It is jointly managed by the AIT Austrian Institute of Technology and Advanced Computing Austria (ACA) and implemented by a consortium spanning the whole of Austria.
The partners in the consortium are: Vienna University of Technology (TU Wien), University of Vienna, University of Natural Resources and Life Sciences, Vienna (BOKU), University of Innsbruck, Graz University of Technology, Johannes Kepler University Linz (JKU), Institute of Science and Technology Austria (ISTA), the Austrian Academy of Sciences (ÖAW), INiTS – Vienna’s High-Tech Incubator and EODC – Earth Observation Data Centre for Water Resources Monitoring GmbH.
AI:AT combines AI-optimised high-performance infrastructure with a physical hub and supports organisations from the private sector, public administration and applied research as a market-neutral, independent access point throughout the AI journey – from the first use case to practical implementation.
https://ai-at.eu/
Funding Acknowledgement
AI Factory Austria AI:AT is co-funded by the European High-Performance Computing Joint Undertaking (JU) under grant agreement No. 101253078. The Joint Undertaking is supported by the European Union’s Horizon Europe programme and by Austria (BMIMI/FFG).