← Neonix Labsopen-source Rust LLM toolstack

RitsuGPT

nanoGPT, in Rust. A small, from-scratch GPT in pure Rust that trains on a single consumer GPU (an NVIDIA GeForce RTX 5060, 8 GB) and runs on your own computer. No Python, no PyTorch — the whole pipeline is plain Rust you can read end to end.

Why Neonix Labs builds it: so that AI knowledge can be truly shared — owned by no one, runnable by anyone. Neonix Labs' goal is a decentralized network where open models and knowledge live in everyone's hands, on their own machines. Small, efficient, open models are how AI stays decentralized.

The bet

AI intelligence is not parameter brute force. An optimized standard architecture plus clean data, and it starts computing on your own machine — no data center required.

What it is, honestly

~16.9M parameters, in the spirit of TinyStories. It learns to write simple, coherent short English stories. It is not a production assistant — no world knowledge, no reasoning, no instruction following. Its value is a clean, hackable, from-scratch stack you can train and check yourself, and it beats the byte-level baseline (bits-per-byte 0.6843 vs 0.805 — lower is better).

Architecture — standard, on purpose

decoder-only TransformerattentionRoPESwiGLURMSNormbyte-level BPE16.9M paramsctx 512

Built with

Rustburncubeclbf16NdArray (CPU)tokenizersMIT

Get it

github.com/NeonixLabs/RitsuGPTopen source · MITtrain it yourself

Pretrained weights

huggingface.co/NeonixLabs/RitsuGPTbpb 0.6843 vs 0.805download · sample on CPU