About us
The people behind BananaMind
BananaMind is a small-language-model project by Banaxi-Tech.
We train compact decoder-only language models from scratch on consumer hardware and release them openly on Hugging Face, together with their tokenizers, benchmark results, and training details. The goal is not to be the largest model — it is to be honest about what a small model trained on a single machine can and cannot do.
Team
Who works on BananaMind
Each profile below links to that person's Hugging Face account, where their public models, datasets, and activity live.
How we work
What we publish, and what we don't claim
Trained from scratch
Every BananaMind 2 model is pretrained from random initialization on a staged data curriculum — no distillation from a larger model, and no continued pretraining of someone else's checkpoint.
Documented runs
Model cards publish token counts, optimizer settings, learning-rate schedules, tokenizer details, and the exact checkpoint each score was measured on, so results can be traced back to a specific run.
Self-reported scores
Benchmark numbers are our own evaluations, labelled as such. They vary with harness version, dtype, and scoring configuration, and preview checkpoints are always marked as previews rather than finished models.
Contact
Get in touch
Questions about the models, benchmark submissions, evaluation results, and commercial licensing enquiries are all welcome. Discussions about a specific model are best opened directly on its Hugging Face repository.