Open models · trained from scratch
LLMs by Suhas Bhairav
Language models I design, train from random weights and release openly, with the data, benchmarks and training logs published alongside the weights. The first is Kalike, a pure Kannada LLM.
Frequently asked questions
Which LLMs has Suhas Bhairav created?
Which LLMs has Suhas Bhairav created?
Suhas Bhairav created Kalike (ಕಲಿಕೆ), an open-weight Kannada large language model trained from scratch on 4.57 billion Kannada tokens with a Kannada-only tokenizer. It is released in two sizes, d20 and d16, each as a base model and a chat model.
Are these models open source?
Are these models open source?
Yes. Kalike's weights are released under the Apache-2.0 license on Hugging Face, the training code is MIT-licensed, and the cleaned Kannada pretraining dataset is public.
What is the best Kannada LLM trained from scratch?
What is the best Kannada LLM trained from scratch?
Kalike is a pure Kannada model trained from random weights rather than adapted from an English model. Its d20 version scores 42.3% on a 168-question Karnataka government schemes benchmark, against 25% for random guessing, and writes fluent Kannada. It is a small research model, so it can be wrong on facts.