A Better Newspaper

## Overview A growing research effort focuses on building large language models specifically designed for languages that are poorly served by mainstream AI systems, aiming to extend AI capability and tool-calling functionality to under-represented linguistic communities. ## Agaram 2.5B A notable example is Agaram 2.5B, described as "the First Tamil-First Language Model Trained from Scratch with Tool Calling." The 2.5-billion-parameter Tamil-English model was trained from scratch on approximately 50 billion tokens, of which 58% was Tamil content, with the remainder consisting of English, math, code, and parallel text (arXiv:2610.18420, October 14). Training reportedly occurred on a single node of eight H100 GPUs over about six days (arXiv:2610.18420, October 14). The model's tokenizer was redesigned to preserve line breaks, addressing a limitation in the developers' earlier model. Following pretraining, supervised fine-tuning was conducted on 134,428 conversations with replies written by an open-weight teacher model, plus additional targeted training rounds, producing a chat model that can respond in Tamil script, romanized Tamil ("Tanglish"), and English, and that supports tool-calling (arXiv:2610.18420, October 14). ## Strategic Significance The project is explicitly framed as "a Playbook for Under-Represented Languages" (arXiv:2610.18420, October 14), suggesting its methodology—efficient small-scale training, tokenizer design, and synthetic SFT data generation—may be replicated for other low-resource languages. This has implications for AI market expansion into non-English-speaking regions, digital inclusion policy, and competitive dynamics among AI developers targeting underserved linguistic markets, which collectively represent billions of potential users.