OP Pohjola and QuTwo take their quantum AI centre into AI model compression

OP Pohjola and the AI company QuTwo are widening their joint research centre, QARC, to compressing AI models with “quantum-inspired” methods. The aim is to let the bank run open AI models at lower cost. The companies give no details of deployment or savings, and the performance figures are QuTwo’s own.

Martti Asikainen | 17.11.2026 | Photo by OP Pohjola

Timo Ritakallio. Photo by OP Pohjola.

Finland’s largest financial services group, OP Pohjola, and the AI company QuTwo said they are widening the work of their joint Quantum AI Research Center (QARC) to model compression. 

The centre explores open language models as an alternative to closed ones, to give the bank more choice in how it uses AI models and manages their cost. 

Compression makes a model a fraction of its original size, so it needs far less computing power, the companies say, while it keeps most of its performance.

Research centre widens its work to model compression

Scaling the use of AI has become one of the biggest new costs for companies, the release says, especially as AI agents take on longer, multi-step tasks. 

It adds that open models are often as capable as, or nearly as capable as, the leading paid services. Compressing them would make them efficient enough to use across the whole business, and not only where the cost is easiest to justify.

Timo Ritakallio, OP Pohjola’s chief executive, said the bank has invested systematically in AI since 2017 and in quantum research since 2019. Through QARC the two lines have come together, he said, which is “a concrete example of what quantum AI can do”. 

QARC was founded in February 2026. At the time, QuTwo said the unit would grow to about 60 specialists by mid-2027. OP Pohjola has more than two million owner-customers and about 15,000 employees.

Shrinking models by 90%

QuTwo’s method has two layers. First come standard techniques for making a model smaller: factorisation, pruning, a step called healing, and quantisation, which stores numbers with fewer bits. 

These lose some accuracy, which QuTwo’s own “quantum-inspired precision-recovery layer” is meant to recover. The company stresses that quantum-inspired methods run on today’s computers, not quantum hardware.

In an earlier announcement, QuTwo said it had tested the method on the LUMI supercomputer, across seven model families and sizes from 70 megabytes to almost 5 terabytes. Models compressed by 74–80% kept 99% of their quality, it said, and models compressed by 90% or more kept 91–95%. 

In one example, a model with 2.4 trillion parameters shrank from 4.89 terabytes to 363 gigabytes, which is 92.5% smaller, while keeping 94.9% of its answer quality. Quality was scored against each full-size original on standard reasoning and knowledge benchmarks.

Quantum-inspired compression of language models is not new. The Spanish company Multiverse Computing published a method called CompactifAI in 2024. 

QuTwo says it has co-founders of the quantum computer maker IQM and the AI company Silo AI on its team, along with more than 50 scientists, and that it has raised a €25m angel round.

White logo of Finnish AI Region (FAIR EDIH). In is written FAIR - FINNISH AI REGION, EDIH
Euroopan unionin osarahoittama logo

Finnish AI Region
2022-2028.
Medialle