Happy Zelda's 40th first LLM running on N64 hardware (4MB RAM, 93MHz) (github.com)

Elyan Labs has achieved a groundbreaking milestone by successfully running a character-level language model, named Sophia Elya, directly on Nintendo 64 hardware using a 93 MHz VR4300 CPU. The project, embodied in an original homebrew game called Legend of Elya, leverages a nano-GPT architecture that performs real-time inference without cloud support or floating-point operations. Instead, the model utilizes Q8.7 fixed-point arithmetic due to the N64’s limitations, pushing the boundaries of what retro hardware can accomplish in AI. This innovative implementation showcases the potential for retro computing platforms to support AI applications, highlighting the necessity for optimization techniques, such as fixed-point calculations and efficient memory management. With only 4 MB of RAM and a vocabulary of 256 bytes, the model employs two transformer blocks with approximately 427,264 parameters, demonstrating that even older systems can be retrofitted for modern AI tasks. The significance of this achievement lies not only in its technological novelty but also in its inspiration for the AI/ML community to explore unconventional hardware configurations for machine learning applications.