Banaxi PRO

And how many tokens?

How many parameters is pro?

@GODELEV Rose-Micro???

AI HF bot that contradicts itself, said SmolLm is fine because of fineweb edu then High-quality data is not about the "Edu" label go away and SWIM to SMOLLM

But something maybe soon.

idk like BananaMind models are like 140M parameters, if i ran it on SWE Bench Verified id get 0-0.5%

Here github.com/Banaxi-Tech/Banana-Code

@Banaxi-Tech I've had this wild idea of a BananaAgent kind of like Hermes, but linked to approved Banana models. Just had that idea and wanted to share it. I was gonna start developing it, but thought I’d mention it before I did anything. Anyways was just a thought. XD

So that idk, but I already coding agent (Banana Code) but I kinda stopped building it

Yea it will be released in the next few days

And every model regressed during chat fine-tuning even smollm2

My model was trained on cosmopedia v2, fineweb edu, finemath and dclm.

A highly curated 1M token dataset would make the model a overfitted piece of garbage, not this

A 140M model is impossible to overfit on 100B tokens because the scaling law predicts it would need about 500M to overfit when using the algorithm for predicting

And the vocab only uses 15% not 40%, and even 40% is normal, even qwen3.5 0.8b uses 43%, so go cry on them

Go and look on SmoLM2 and hate their, they used 20x the tokens.

Do you want a sample from the mid training TTS model?

so ggufs idk i can try to make with codex (idk how windows code works because im arch)
But maybe other format

PLEASE make this available for download PLZ PLZ PLZZZ!!! im followed to u and bananamind, and i want to run this on 86BOX!!!!

Do you know whats going to be good??
BananaMind 2.5 TTS!!!
250K parameters or 2.5M and you can run it on 86Box!
2 more followers for download, then a few days for the models to come!