Roadmap
OpenCerebral publishes what it has finished rather than what it intends, so this page is short and the dates are soft. Nothing here is a commitment.
In progress
- Boris-1.3-250M. Extending the continued-pretraining recipe to the 250M checkpoint. The open question is whether the FineWeb-Edu-first ordering that avoided the ARC regression at 125M holds at the larger size.
- littlerock scaling series. More points between 1M and 75M, trained identically, to get a clean picture of where each benchmark leaves the chance floor.
Wanted
- Independent evaluation. Private variety sets written by someone other than us. See Get Involved.
- Reproductions. Someone else running a published recipe on their own RTX 3060 and reporting whether the numbers match.
Under consideration
- A held-out evaluation suite that is not authored by the person training the models — the direct lesson of the arithmax result.
- Longer context for the Boris family. 1024 tokens is the current limit.
- Publishing the training and evaluation code alongside the weights.
Larger models
The published range so far runs from 1M to 250M parameters because that is what the hardware on hand finishes in hours to days. It is a description of the current compute budget, not a ceiling we have chosen. Bigger models are wanted and intended; what stands between here and there is GPU time. If you can help with that — compute credits, or a card sitting idle — see Get Involved.
What will not change as the models grow is the method: the full recipe, the wall-clock figures, the failed passes, and permissive licensing wherever the training data allows it.
Hosted endpoints
We do not run an inference API ourselves — the models are small enough to run locally on a CPU, so it has not been the best use of the budget. Anyone who wants to host one is welcome to, and we would be glad to link it. Apache 2.0 with no field-of-use restriction means you need no permission from us. Tell us at oc@kunix.org and we will point people at it.