Project Titania
A from-scratch large language model stack, covering the model, an instruction set, a compiler, a simulator, and a GPU design, built small enough for one person to read end to end.
Project Titania is a complete large language model system built from scratch at every layer: the model itself, a custom instruction set architecture, a compiler that lowers model kernels to that instruction set, a simulator that executes the instructions, and a planned GPU implementation. It currently runs Qwen3-0.6B, a small decoder-only transformer, with its kernels compiled to the Titania instruction set and executed today by an instruction-set simulator; an RTL GPU, an FPGA, and eventual silicon are planned next steps.
The instruction-set simulator is the single source of truth for what each instruction does, and the hardware is required to reproduce its results bit for bit. Each layer is deliberately kept small and documented so that one person can read, understand, and implement all of it, rather than building on an existing GPU, instruction set, or compiler as a black box.
It installs with cargo and downloads its model on first use; running it lets the model use a bash tool to inspect files and the host system without confirmation.