Bootgpt Fits a Character-Level Language Model Into a 512-Byte Boot Sector Plus 23 Sectors of Weights

A GitHub project called bootgpt boots a PC straight into a small neural network that writes fake Linux boot logs. Its author, kerlenton, published it under the MIT licence, and it was posted to Hacker News as a Show HN.

The README describes it as "A neural network that boots your PC." The boot sector is 512 bytes of hand-written assembly. The model is an 11.5 KB int8 file stored in the 23 sectors that follow.

What runs at boot

The boot sector loads the weights through BIOS interrupt 13h and prints through interrupt 10h. It then runs inference in 16-bit real mode. The README says there is "no OS, no C runtime, nothing."

The architecture is character embeddings, one hidden layer and a softmax output. The example output uses 8 characters of context. Weights are int8 with power-of-two scales. The README gives no parameter count.

The README says it runs on a 386. It can be tried in QEMU or written to a USB stick with dd and booted on a BIOS machine. Any keypress reboots it.

Training data and output quality

A script called corpus_gen.py generates about 3 MB of fake Linux dmesg logs, and the model learns to continue them. The author describes the result this way: "The machine wakes up and dreams a Linux boot log: timestamps, PCI bars, ACPI events for hardware that never existed."

The sample output in the README is rough, with garbled text and invented hardware. The README says the Python simulation and the boot sector output match "char for char." It gives no timing figures.

Project	Code size	Model	Notes
bootgpt	512-byte boot sector	11.5 KB int8, 23 sectors	One hidden layer, 8-character context
sectorllm	1,189 bytes	stories260K, 260K parameters	5 layers, 8 attention heads

The transformer comparison

The README compares bootgpt with other boot-sector projects, including sectorllm. That project's README, by the user rdmsr, calls it "The world's smallest llama2 inference engine" and says it "fits in 1189 bytes of x86 real mode assembly."

Sectorllm runs stories260K, a children's-story model with 260K parameters, 5 layers, 8 attention heads and a 512-token vocabulary. It switches the CPU into unreal mode so the KV cache fits. Its README says the model architecture and prompt are hardcoded, sampling is greedy argmax only, and the code is golfed so performance and precision are not optimal. It shows no licence.

The two projects make different trade-offs. Bootgpt keeps the code to one sector and uses a far simpler network. Sectorllm needs more than one sector for a full transformer pass.

What it is and is not

Bootgpt is trained for one task and is not a general text generator, per its README. It is a demonstration of how little code inference needs, not a usable model.

For other open-source projects that spread on GitHub, see Lightpanda 1.0 and claude-mem. The README names the corpus generator and the QEMU route, so anyone can try the 11.5 KB model for themselves.