The architecture, the C99 engine, the ternary export format and a 944,000-parameter TinyStories checkpoint are public under Apache-2.0. This page says exactly what is in the box and what is not, because "open source" on an AI project can mean anything from a full training path to a weights download with a licence attached.
What is public
| Artifact | Where | What it contains |
|---|---|---|
| Source | github.com/TilelliLab/atome-lm | The three-pathway architecture, the zero-heap C99 inference engine, the ATOME01 packing format, the parity tests, the RAM measurement script. |
| Weights | huggingface.co/TilelliLab/atome-lm | The trained checkpoints, including the 944,000-parameter TinyStories model. |
| Citation | 10.5281/zenodo.20518644 | A citable DOI for the release. |
| Licence | Apache-2.0 | Commercial use permitted, patent grant included, attribution required. |
What is not public, and why
- Training scripts for the specific classifier prototypes. The engine path for running a classification head is public; the data generators and training recipes for the three internal prototypes are not.
- The elaborated research variants. Extended pathways, retrieval and memory, multi-bank weight schemes and an internal regression-prevention gate stay internal.
- The EDGE benchmark harness. The sensing line and its UCI HAR harness are part of a pilot engagement rather than the public kit. If you want to run it against your own data, ask — that is a normal request and the answer is usually yes.
Reproducing the published numbers
Every figure on this site that concerns the public kit comes from a script in it:
- Parity —
tests/test_parity_with_c.py,tests/test_parity_multitoken.py - RAM per model size —
python3 scripts/measure_ram.py --markdown - Engine size —
arm-none-eabi-sizeon the cross-compiled object
The full reading of the perplexity results, including the ones where the
architecture loses, is in HONEST_RESULTS.md in the repository.
Using it in a product
Apache-2.0 means you can ship it commercially without asking. What people generally want help with is not the licence but the bring-up: getting a model of the right size onto a specific part, secure boot and signed model loading, per-platform hardening, and the evidence pack a certification body will ask for. That is what the services page is about.
Citing it
If Atome LM is useful in academic work, cite the Zenodo DOI 10.5281/zenodo.20518644. If you benchmark it and it loses, publish that too — we do, on the limitations page, and a negative result about a small model is worth more to the field than another positive one about a large one.