Launching today

Kalike
A pure Kannada language model (435M), trained from scratch
2 followers
A pure Kannada language model (435M), trained from scratch
2 followers
Kalike has a clear launch story: it’s an open‑weight Kannada language model trained from random weights, with Kannada-only data and tokenizer. The Hugging Face page reports a 435M-parameter core model, 4.57B pretraining tokens, and a smaller chat-tuned version.



Free
Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted
Hi Product Hunt! I’m Suhas, an applied AI engineer, and I built Kalike to explore what it takes to train a small language model in Kannada from scratch.
Kalike’s tokenizer, pretraining data, and chat data are Kannada-only. It has a base model for continuing Kannada text and a chat-tuned version. The weights, training details, evaluation results, and examples are available on Hugging Face.
This is an early research model, not a production assistant. It can produce fluent-looking Kannada but may invent facts, and it has not had safety tuning.
I’d especially value feedback from Kannada speakers and language-model builders: Which Kannada tasks should I evaluate next, and what would make a small Kannada model more useful to you?
Read more about it on https://suhasbhairav.com/llms/kalike