The hub
Models, with their receipts attached
A model page should tell you what the thing measures, how it was measured, and which numbers cannot bear weight. Ours starts with the model we trained on our own bench, and it is built to hold yours.
- Open weights
HomeBrewedLabs
Metallum-1B
A 1,002M-parameter ML / LLM-engineering specialist, pretrained from random initialization on one consumer GPU.
- text-generation
- ml-engineering
- 1B
- trained-from-scratch
- long-context
- fill-in-the-middle
- single-GPU
- Parameters
- 1,002M
- Context
- 2048 tokens
- Licence
- Apache 2.0
- Updated
- 2026-08-07
- Open weights
HomeBrewedLabs
Metallum-550M
The 550.3M-parameter first Metallum: an archival first-generation ML / LLM-engineering specialist trained from scratch on one consumer GPU.
- text-generation
- ml-engineering
- 550M
- trained-from-scratch
- retrieval
- custom-code
- single-GPU
- archival
- Parameters
- 550.3M
- Context
- 2048 tokens
- Licence
- Apache 2.0
- Updated
- 2026-09-01
Publish your model here
The hub is being built to hold more than ours. Bring a model you trained, whether from scratch, fine-tuned, or distilled, and serve it from infrastructure sized for one-card budgets instead of renting a hyperscaler. Uploads open in waves alongside the hosting platform.
- Your weights, your licence, your call on who downloads them
- Evidence classes on every published metric, same rules we hold ourselves to
- Private and gated models supported, not just open ones