Gargi-M1 is a 110M-parameter Malayalam language model, trained from Malayalam rather than translated into it. Both checkpoints are open. This is the short account of what it is and what it still cannot do.
Gargi-M1 is public today. The base and instruct checkpoints and the tokenizer are on Hugging Face under Apache 2.0, and you can try it in chat right now — switch between the two checkpoints in the header and watch them disagree.
| Parameters | 110M — 12 layers, 12 heads, 768 hidden |
| Context | 512 tokens |
| Vocabulary | 32,000 byte-level BPE, Malayalam only |
| Training | 1.39B tokens, held-out loss 1.189 |
| Tokenizer fertility | 1.36 chars/token |
| Licence | Apache 2.0 |
A decoder-only transformer in the GPT-2 shape — deliberately conventional. At 110M, the two things that move the number are the tokenizer and the data, not the architecture, so that is where the work went.
Malayalam is agglutinative, and most multilingual tokenizers split its conjunct forms three or four ways. This one is trained on Malayalam alone and averages 1.36 chars/token. That is not a benchmark flourish — it is the context window. Fewer tokens per word means more real Malayalam inside the same 512 tokens, and more documents seen per compute hour. At this size the tokenizer is the budget.
Gargi-M1 is a research preview. 512 tokens of context is short, and 110M parameters holds limited world knowledge — it will state things confidently that are not true. Instruction following is early, and the corpus is drawn largely from written, formal sources, which is not how most of Kerala actually speaks. None of that is hidden behind a waitlist: the failures are in the same chat as the successes.
Gargi-M1 is the first step of something larger — an open language layer for Indian languages, not a single model — and that is not a thing one person finishes. So this release is also a call for people. If any of these are you, I would like to hear from you:
You do not need a machine-learning background to help — some of the most useful work right now is reading model output and saying precisely where it is wrong. The weights, the tokenizer and the evaluation code are open and stay open. Take them apart, then email gishnum.work@gmail.com or leave your address below.