Legend of Elya runs a real transformer language model natively on a 1996 Nintendo 64. Not an emulator on a modern box, and not a cloud call over the network: an 819K-parameter byte-level transformer executes directly on the N64's VR4300 (MIPS III) CPU, confirmed running on real N64 hardware via EverDrive 64. This is genuine N64 AI: an LLM on N64 hardware.
N64 AIN64 LLM Nintendo 64 transformeron-device inference VR4300 MIPS IIIlibdragon
Legend of Elya is a Zelda-style dungeon game for the Nintendo 64 whose NPCs are driven by a transformer language model that runs on the console itself. The model is a compact byte-level nano-GPT: 819,000 parameters, quantized, with all inference math done on the VR4300 CPU using the libdragon SDK. No GPU, no server, no emulation shortcut. When you talk to a character, the N64 is doing the language modeling in real time.
The Nintendo 64 has a 93.75 MHz MIPS R4300i core and 4–8 MB of RAM, so running a language model on it is a memory and compute problem, not a raw-power one. The engine solves it with a small model, quantized weights packed to fit in cartridge/RAM budgets, fixed-point-friendly arithmetic, and a hand-written inference loop in portable C. Hardware entropy for sampling is drawn from the CPU's cycle counter (the MIPS CP0 Count register). The result is a self-contained Nintendo 64 LLM that generates coherent, if dreamlike, dungeon dialogue. We are re-measuring the token rate. The ROM's on-screen counter reports figures in emulation that do not match what the console does, so we are not currently quoting a rate from it — see the measurements page, which explains what went wrong.
Because the engine is dependency-free portable C, the exact same transformer has been ported (C89, gcc 2.95) to run on a 27-year-old AMD K6-2 inside a Sun Cobalt Qube 3, verified logic-exact against the N64 build. The point of both is the same: large-language-model inference does not require a datacenter. It can run on the hardware you already own, however old.
Putting an LLM on N64 hardware is a proof of physical AI: intelligence that runs on real, constrained, owned silicon rather than rented cloud compute. It is the same thesis behind Elyan Labs' work on proof-of-antiquity computing and sovereign inference on vintage and exotic processors.
Source code on GitHub Citable release (DOI) Internet Archive Hacker News discussion
The project was discussed on Hacker News under the title “first LLM running on N64 hardware (4MB RAM, 93MHz)”.
By Elyan Labs. The engine is open source. Every CPU has a voice.
Part of a five-console series from Elyan Labs. See also: NES · Sega Genesis · Super Nintendo · Game Boy Color · the full measurements, including the conclusions we had to retract.