As artificial intelligence continues to evolve, well-known models like ChatGPT, Claude, and Gemini still use the Transformer architecture that Google introduced back in 2017. However, a research lab in Stockholm called Farang is taking a new path by developing language models that prioritize internal reasoning. This means their models aim to fully grasp what they want to say before putting it into words, mirroring how humans think before they communicate.
Recently, Farang secured €1.5 million in seed funding, giving the company a valuation of €10 million post-investment. The funding round was led by Voima Ventures and Amadeus APEX Technology Fund, along with support from several well-known Nordic deep-tech entrepreneurs and angel investors.
The technology from Farang aims to significantly lower the costs of training and operating AI models by reducing computational needs by 25 times compared to existing models like ChatGPT.
Rethinking Language Technology
Farang’s founder, Emil Romanus, has extensive engineering experience and a successful background in startups. Before starting Farang, he co-founded Present Communications (now known as TalkTastic), which raised $8 million to develop AI technology focused on video.
Romanus shared his perspective with us, explaining that he started Farang after noticing the limitations of the Transformer architecture. Originally created for translating languages, this structure wasn’t designed to handle complex reasoning or problem-solving tasks.
“By chance, these models began to exhibit ’emergent abilities’ only at large sizes,” Romanus noted. “However, these abilities should ideally show up earlier if the architecture is designed with specific tasks in mind.”
Currently, Farang operates with a small team of five, all working remotely from Stockholm and nearby areas.
Understanding Farang’s Technology
One of Farang’s main focuses is on the React development community, which includes about 40% of professional programmers worldwide. Romanus points out that existing large language models often struggle during iterative software development. They can produce repetitive or conflicting code, resulting in bugs and making future maintenance difficult.
“React poses a significant challenge to leading models like Claude 4 and GPT-5. When products move beyond simple questions to iterative development, these models struggle to track and update components. This can lead to redundancy and mismatched code, which complicates maintenance.” He added, “By concentrating on React for one of our first models, we aim to offer a tailored solution for a large user base, creating real business value instead of just flashy technology.”
Farang sees major competitors in Google, OpenAI, and Anthropic. Over the next three to five years, Romanus plans to showcase Farang’s technology on a larger scale and aim to outperform these industry leaders in key performance areas, focusing on challenges that current models handle poorly.
Ultimately, Romanus aims to make Farang the preferred choice for specialized, enterprise-ready AI, competing directly with OpenAI for user adoption.
