Researchers have successfully executed a large language model on hardware from the late 1990s, proving that modern artificial intelligence does not strictly require high-end graphics cards or massive memory banks. By utilizing a computer equipped with a 350 MHz Pentium II processor and 128 MB of RAM, the team demonstrated that AI can function on systems that have been obsolete for decades.
The experiment, highlighted by TechSpot, utilized a model based on Llama 2 running on Windows 98. This feat was made possible through the combined efforts of Andrej Karpathy and the team at EXO Labs, an organization dedicated to expanding access to artificial intelligence technology.
BitNet architecture enables efficiency
The breakthrough centers on BitNet, an architecture that utilizes ternary weights. By compressing 7-billion-parameter models into a footprint of just 1.38 GB, the architecture allows for significant model reduction without sacrificing core functionality. This technical shift prioritizes CPU processing power over the specialized hardware usually required for machine learning tasks.
According to EXO Labs, this optimized approach achieved performance speeds of approximately 39 tokens per second. The developers noted that this method could improve overall efficiency by up to 50 percent. By shifting the workload away from expensive, power-hungry infrastructure, the researchers aim to democratize access to AI tools.
This development suggests a future where large-scale models operate reliably on standard consumer CPUs at speeds comparable to human reading. As the technology evolves, it could eliminate the current reliance on high-performance hardware, allowing AI integration into a wider array of legacy devices and everyday consumer products.