Models built from scratch by two people on consumer GPUs. We make random stuff, test out architectures and push limits.
This org is simply where we host our models. Everything here is an experiment: expect research code, honest model cards, and results that are sometimes "it didn't work".
| Model | By | What it is | Status |
|---|---|---|---|
| Meridian‑Mini | hashtagg1 | 236M-parameter MoE with hybrid KDA/MLA attention, latent experts, manifold-constrained hyper-connections, attention residuals and Engram n-gram memory | Training |
| Meridian‑Tiny | hashtagg1 | 11.8M-parameter smoke test of the Meridian architecture, with component ablations | Released |
| Antimatter | CraftingScrpt | Low (million-parameter tier) multimodal MoE inspired by Tiny LLM | Training |
Each model card covers the architecture, training data, training setup and limitations in detail.
This org is maintained by the two of us:
Questions, bug reports or ideas are welcome in the Community tab of any model.