A 99.94M parameter bilingual decoder-only language model trained from scratch on IndicCorpV2, Bangla Wikipedia, and educational web text. Features RoPE embeddings, RMSNorm, and SwiGLU non-linearities for robust low-latency inference.
Kotha-S1
Explore Kotha Lab Foundation Models
Browse open model weights, training checkpoints, and reproduction code on our Hugging Face hub.