Farang, a newly launched AI research lab in Stockholm, has raised €1.5 million in seed funding to develop what it says is a new foundational AI architecture designed to outperform existing transformer-based models while using dramatically fewer computational resources.
The round was led by Voima Ventures and the Amadeus APEX Technology Fund, with participation from angel investors including Tero Ojanperä (Silo AI co-founder), Nilay Oza, and Niraj Aswani (Klevu founders).
Rethinking large language models
Founded in 2025 by Emil Romanus, an AI engineer with 20+ years of experience in search and machine learning, Farang is developing models that generate full responses conceptually before translating them into text — a shift from the word-by-word prediction used by transformers.
This architecture, according to the company, results in more coherent outputs while cutting computational requirements by 25x.
Farang’s initial focus is on specialised domains, including programming assistants for frameworks like React, as well as medical AI tools.
“We’re not just building another layer on top of transformers — we’ve created a new architecture from the ground up,” said Romanus. “It allows organisations to deploy specialised models entirely on-premises with full privacy controls.”
Advertisement
Privacy-first AI for enterprises
By enabling customers in healthcare, finance, and legal services to train and run models in-house, Farang addresses growing concerns around data sovereignty and regulatory compliance.
The five-person team plans to use the new funding to scale proof-of-concept models, expand compute capacity, and target developers and AI enthusiasts via a waitlist program.
Global ambitions from Europe
Backers see Farang as part of Europe’s bid to compete in foundational AI research.
“Farang is a paradigm shift — efficient, specialised, and enterprise-grade from day one,” said Inka Mero, Managing Partner at Voima Ventures.
Long-term, the startup aims to challenge industry leaders like OpenAI by first proving its architecture in niche verticals before expanding into general-purpose AI.
