A small independent team has shown that a serious language model for an Indian language does not have to start with a nine-figure budget. Murai Labs has released Agaram 2.5B (அகரம்), an open Tamil–English language model trained from scratch, along with its code, under the permissive Apache-2.0 licence. The pretraining run cost roughly USD 4,765 in rented compute.

The model chats in Tamil script, in romanised Tamil (the “Tanglish” that dominates everyday messaging) and in English, and it can call external tools. The base model, a chat model, GGUF files for running it locally on llama.cpp and Ollama, and the training code were published on 27 October 2026, with a technical report on arXiv.

Why build from scratch

The project was led by Ramchand Kumaresan, who began collecting data in December 2025 and was joined by two volunteers in mid-2026. The team’s core argument is about tokenisation, the way a model breaks text into pieces. Most popular vocabularies are learned mostly from English text, so Tamil gets chopped into many fragments. On Murai Labs’ samples, Qwen2.5’s tokenizer needs about 9.4 tokens for an average Tamil word, compared with 1.35 for an English word, which makes Tamil roughly seven times more expensive to process.

Bolting new Tamil tokens onto an existing model leaves them untrained, so the team designed a 64,000-token vocabulary with Tamil first. It averages 1.55 tokens per Tamil word, keeps Tamil vowel signs attached to their letters and reproduces text exactly, including line breaks. Training from scratch also meant the team knew every document the model saw: before reporting any benchmark, it removed 42,001 documents that overlapped with 20 evaluation sets.

One run on one rented node

Agaram 2.5B is a dense decoder-only Transformer with about 2.53 billion parameters. It was pretrained on 50 billion tokens, 58% of them Tamil and the rest English, maths, code and parallel text, in a single run on eight NVIDIA H100 GPUs that took about six days and finished on 8 October 2026. Compute came partly through credits from the NVIDIA Inception programme and JarvisLabs. Across both of its models, the team puts the full project cost at about USD 13,900.

That second model exists because the first one failed. An earlier 1.5B-parameter version, trained in July, was never released: a tokenizer bug erased every line break, so the model could not learn lists, verse, code or even where a chat turn ends. The team said lessons from that failure, from the data mix to how checkpoints are safeguarded, went directly into Agaram 2.5B.

Strong on translation, weak on facts

Murai Labs is unusually candid about the model’s limits. On Tamil–English translation it beat Sarvam-1, a 2B multilingual base model, in all six directions measured. On reading comprehension and Tamil MMLU, however, it scores near chance and well below Sarvam-1, which was trained on around four trillion tokens against Agaram’s 50 billion.

On 50 everyday prompts written by a native speaker, an AI judge rated about half of the released chat model’s replies usable, with most failures being wrong facts. On 25 tool-use prompts it picked the right tool every time. The team describes Agaram as a research model that has not been red-teamed and should not be used where a wrong answer could cause harm without human review.

Why it matters

India’s AI conversation is often framed around large national foundation models and GPU tenders. Agaram is a reminder that language-specific, open and well-documented models can also come from small teams on modest budgets, and that the real bottleneck for many Indian languages is clean training text. Murai Labs says much printed Tamil exists only as scanned images and that finding more text from books, archives and public institutions is where the next effort should go.

For enterprises and public-sector teams building citizen-facing services in regional languages, an openly licensed model that can run locally and call tools offers a useful base to experiment with, provided they add their own checks on accuracy and safety.

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!