An open Tamil–English language model, trained from scratch. Agaram 2.5B chats in Tamil script, in romanized Tamil (Tanglish) and in English, and it can call tools. This program was called TamilLM until the model was named.
Try the previewPublic chat demo on Hugging Face, open until 18 October 2026.
Tamil has tens of millions of speakers and a written literature more than two thousand years old. Open language models still serve it poorly.
Part of the reason is the tokenizer. Most vocabularies are learned mainly from English. On our samples, Qwen2.5's tokenizer splits an average Tamil word into 9.4 tokens and an English word into 1.35, so the same number of words costs about seven times as many tokens in Tamil. Adding Tamil tokens to an existing model leaves them untrained. Training from scratch let us build the tokenizer for Tamil first.
The name came from a public poll. Agaram (அகரம்) is the letter அ, the first letter of the Tamil alphabet.
A 64,000-token byte-level BPE trained on a sample that was 58% Tamil. It averages 1.55 tokens per Tamil word, keeps Tamil vowel signs attached to their letters, and reproduces text exactly, line breaks and indentation included.
We tested hybrid architectures with Mamba layers at about 91M parameters, three seeds each. None beat the standard Transformer on held-out Tamil text, so Agaram is a dense decoder-only Transformer. That test stopped at about 100M parameters; it does not show that hybrids cannot win at larger scale.
Training from scratch means we know every document the model saw. Before reporting benchmark scores we scanned the training data against 20 evaluation sets and removed 42,001 overlapping documents.