Scientists at Pathway have introduced a new AI model called BDH-CQ, which utilizes an entirely unconventional approach to reasoning. According to researchers, this new vector-based technology is significantly cheaper than standard methods and could signal the beginning of a "post-transformer" era in artificial intelligence.
To test the efficiency of the new system, experts used the well-known ARC-AGI benchmark created in 2019, which measures the non-verbal reasoning level of AI systems. BDH-CQ was able to correctly solve around 30 percent of the tasks, which is quite an impressive result considering its small size. Although some larger models achieve higher scores, their operating costs are immense. For example, OpenAI's GPT 5.6 Luna (Low) model recorded slightly higher accuracy, but its cost was roughly 11 times higher.
The main reason for this difference is that the new model has only 150 million parameters, whereas today's most powerful systems have tens to hundreds of billions of parameters. Thanks to its small size, the model learns faster and requires far less electricity and computing power.
Conventional models write out their chain of thought word by word when solving problems, which strains memory and makes the process expensive. BDH-CQ instead uses special numerical arrays and internal recurring cycles, allowing it to solve complex abstract problems without logging additional text.
The results of the new architecture have already been verified and confirmed by independent researchers. Pathway plans to scale this technology up to 600 billion parameters in the future and apply it in cybersecurity and industrial sectors.
The results recorded by the BDH-CQ model show that progress in the field of artificial intelligence can occur not only by endlessly increasing model sizes, but also by radically changing their architecture. If this "post-transformer" approach gains widespread adoption, it could significantly reduce the enormous costs of maintaining and deploying AI systems.
However, it is still too early to claim that the technology is fully ready to replace existing standards. The model has only been tested on specific types of reasoning tasks, and Pathway still needs to prove that this system works just as effectively with large volumes of text and standard chatbots.
