Frequently Asked Questions
What happened to NMAI?
Nothing, other than the name. New Millennium Artificial Intelligence is
now OpenCerebral. Same models, same maintainers, same parent project. Model
cards and repository paths that still say NMAI or
KSP-NMAI refer to OpenCerebral.
How are these models good enough to be useful?
For most purposes people mean by "useful," they are not — yet. A 125M-parameter model scoring 39.38 on our six-benchmark mean is not going to answer your questions correctly. What these models are good for is research on scaling and data mixtures, education, offline and embedded experiments, and as small finetuning bases. We would rather say that plainly than oversell them.
Why one GPU?
Partly because it is the hardware we have, and partly because it turns out to be a good property: a published recipe that requires a cluster is not one most people can check. Every number on this site came off a single RTX 3060 and is reproducible by anyone with the same card and some patience. We intend to keep that reproducibility as the models get bigger, but we are not going to pretend the current size range is a deliberate ceiling.
Which model should I download?
For conversation, Boris-1.3-125M-Instruct. For text continuation or finetuning, Boris-1.3-125M. For CPU-only use, a GGUF build. For research at the smallest scale, littlerock-1M.
What is the difference between Boris and Boris 1.3?
Boris 1.3 is the original checkpoint plus continued pretraining passes that mix in DCLM-baseline. It scores better on every benchmark mean we track. Use 1.3 unless you specifically want the original for comparison.
Why publish a model that failed?
Because how it failed is the useful part. A finetune that passes an independent memorization test and then scores below chance on rephrased questions is a clean demonstration that held-out splits you wrote yourself cannot detect distribution-level specialization. That is worth more than another incremental win.
Can I use these commercially?
Yes. Apache 2.0, no acceptable-use rider, no field-of-use restriction. See Licenses.
Are they safe?
No. They have received no alignment or safety tuning and will produce inaccurate, inconsistent, or offensive text. Do not deploy them unsupervised or treat their output as factual.
Is there an API or a hosted demo?
Not one we run. The models are small enough to run locally on a CPU — see Usage — so hosting has not been the best use of our budget. If someone else wants to stand one up, good: Apache 2.0 carries no field-of-use restriction, you need no permission from us, and we will link it from this site. Let us know at oc@kunix.org.
Will there be bigger models?
Yes, as compute allows. The models are small because consumer GPUs are what we have to train on, not because we think small is where this should stop. A 2.5B-token run at 125M is about a day and a half on an RTX 3060; the same recipe at a size worth calling large is not something one desktop card finishes. That is a budget problem, and it is the one place where outside help changes what this project can publish — see Get Involved.