Agaram 2.5B

Agaram 2.5B

Chats in Tamil script, romanized Tamil and English, and translates between Tamil and English. A 2.5B model built from scratch by Murai Labs, for Tamil speakers, researchers and developers.

Made by murai-labs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Agaram 2.5B is a chat model with 2.5 billion parameters for Tamil and English. Murai Labs and Ramchand Kumaresan built it from scratch, and the card says it uses no weights from other models. It is a non-profit research model for chatting in Tamil script, romanized Tamil and English, and for translating between Tamil and English. This repository holds the chat version, made from the pretrained base with one supervised fine-tuning run.

Who it suits: It suits Tamil speakers and builders of Tamil tools who want a small model that handles Tamil, romanized Tamil and English in one place.

Industry | Who uses it | How they use it | Example
Education | Tamil learners and teachers | Chat and practise in Tamil or romanized Tamil | Ask a question in romanized Tamil and read the reply
Language tech | Developers | Build Tamil and English chat or translation features | Send a Tamil sentence and request an English version
Research | Language researchers | Study a model trained from scratch with mostly Tamil data | Test how it handles Tamil prompts

If you need other languages or long inputs, you can skip it, since the card lists only Tamil and English and a context of up to 8,192 tokens.

Where to use it: You run it in Python with the transformers library, on Windows, macOS or Linux. The card also points to an online chat demo and a chat command line tool in its GitHub project.

• Python scripts that use transformers 5.x
• Windows, macOS or Linux with the huggingface_hub package
• The online chat demo linked from the card

Use case | Who | What they get back
Tamil chat | Learners and users | Replies in Tamil, romanized Tamil or English
Translation | Developers | Tamil to English or English to Tamil text
Tool calls | Developers | A tool request in a set format, with results sent back as messages

How to set it up: You need Python, the transformers library at version 5.x, torch, and about 5.1 GB of space for the BF16 weights. The card says the 5.x format is needed to store the rotary base.

• Python on Windows, macOS or Linux
• transformers 5.x and torch
• About 5.1 GB of free space

1. Install the Hugging Face hub package.
2. Download the model files.
3. Run the card's Python example.

pip install -U huggingface_hub
hf download murai-labs/agaram-2.5b-instruct

How to use it: Load the tokenizer and model, build a chat message, and generate with greedy decoding, up to 512 new tokens. No system prompt is needed. The stop token has id 2.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "murai-labs/agaram-2.5b-instruct"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Pencil-la thappa ezhudhiten. Adhai azhikka edhai use pannalaam?"}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)["input_ids"]
out = model.generate(ids.to(model.device), max_new_tokens=512, do_sample=False, eos_token_id=2)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Tool | What it does
AutoTokenizer.from_pretrained | Loads the tokenizer made for this model
AutoModelForCausalLM.from_pretrained | Loads the weights in bfloat16
apply_chat_template | Builds the chat prompt, and takes a tools list for tool calls
generate | Produces the reply with greedy decoding

Example: You send the card's romanized Tamil question about what to use to erase a pencil mistake. The model returns its answer as text, which the script decodes and prints.

For newcomers and experts: If you are new to this, it works like a chat assistant that understands Tamil, including Tamil typed in English letters. A first small step is to run the card's example script and then swap in your own question.

If you know the field, it is a 28-layer dense Transformer that normalizes before each block, with RMSNorm, SwiGLU layers and rotary position encoding. Its config borrows the Llama class name because that code can load the design. It has a byte-level BPE tokenizer with 64,000 tokens trained for it, about 50B pretraining tokens, and a context of 4,096 tokens, raised to 8,192 in the decay stage and fine-tuning.

Tamil-heavy pretraining data
Base model trained from scratch
One supervised fine-tuning run
Chat model for Tamil and English

Good to know: Part of the chat data was written by an open-weight teacher model, and the card lists exceptions under its disclosures section. Its chat command line tool adds a crisis-resource footer.

Title: Pretraining mix by share of tokens
Tamil: 58.2 percent
English: 19.8 percent
Math: 10 percent
Code: 7 percent
Instruction: 3 percent
Parallel text: 2 percent

• License: Apache-2.0, with the upstream data terms described in the card
• Cost to run: no API key or paid service is named, and the weights are about 5.1 GB in BF16
• Data it reaches: the download comes from Hugging Face, and the model works on the text you give it on your own machine
• Not for: languages beyond Tamil and English, or inputs beyond its 8,192-token context

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 2.5B
• Weights: about 5.1 GB in BF16; about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,296
• Likes: 11
• License: apache-2.0
• Languages: ta, en
• Based on: murai-labs/agaram-2.5b
• Quantized versions: 1
• Tags: transformers, safetensors, llama, text-generation, tamil, tanglish, agaram, chatml, from-scratch, conversational

Links

• Model: https://huggingface.co/murai-labs/agaram-2.5b-instruct
• Files: https://huggingface.co/murai-labs/agaram-2.5b-instruct/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download murai-labs/agaram-2.5b-instruct

Then follow the usage in its model card. Nothing in this kit runs by itself.