⚠️ Research preview - everything here is still under test

All models in this organisation are experimental checkpoints, published so people can look inside them. None of them is finished, and none of them is reliable.

• The best published checkpoint reaches 37.37% held-out next-token accuracy (by position). That means most of its next-word guesses are wrong.
• Its context window is 32 tokens. It cannot hold a conversation, follow instructions, or answer questions.
• In our own published tests, an attention block lost to having no attention at all, and our own architecture lost to a plain one. Both are in the repo with their log lines.

Please do not use any of this in a product, a service, or anything a person depends on. The downloads are for inspection, testing and reproduction. If you find something broken, that is expected - tell us and we will publish the fix.

Welcome to Blueprint AI HQ

We build machine learning models you can look inside.

GlassNet is our first project: a language model written entirely from scratch - our own tensors, our own backward pass, our own optimiser, our own KV cache. No PyTorch, no transformers, no autograd library, no pretrained weights. The whole architecture is about 7,000 lines of readable NumPy.

Our Main Focus Areas

• Interpretable language models - every middle neuron carries a human-readable name, and a neuron whose behaviour does not match its name is dropped, never relabelled
• From-scratch infrastructure - tensors, gradients, optimiser and KV cache written by hand and checked numerically
• Honest measurement - every claim names the run it came from, including the experiments that failed
• Shipping real weights - checkpoints published in standard containers so anyone can test them, not just read about them

Our Models

Context window is 32 tokens - about one short sentence. That is a hard limit of the current setup, and we state it up front because overstating a model is the default failure of this field.

How We Work

1. No number without a run behind it. Every figure names the file it came from.
2. Failed experiments are published. Our own architecture lost to the simple one, and that loss is on record.
3. If a claim cannot be audited, it is dropped rather than softened.
4. We say how bad it is. GlassNet is not an assistant yet, and it scores 37.37% held-out.

Code And Results

Full source, tests, charts and every run log including the failures: https://github.com/blueprint-ai-hq-git/glassnet

Testing Our Work

Run the gradient checks, reproduce a number, or try to catch an explanation lying - report anything you find, including results worse than ours. We publish those next to our own.

Start with docs/TESTING.md in the repository.

Contact Us

Email: blueprintai.help1@gmail.com

Open a discussion on any model page, or an issue on the repository. For anything security or privacy related, email first rather than posting publicly.