An honest look at the competition, and why there is still room for you.
The short version¶
There is real activity around Indic languages, and some dedicated Sanskrit work. But there is no mature, high-performing, specialised classical Sanskrit model that scholars and students actually use for deep work.
The gap for a strong, accessible, local-first model is real.
The large efforts¶
Bhashini — the Indian government’s national language technology mission. Covers 22-plus scheduled languages with translation, voice, and multilingual models. Sanskrit gets some support inside the broader stack, but deep classical work is not the focus.
AI4Bharat (IIT Madras) and related academic groups — strong translation models including Sanskrit, multilingual Indic models, benchmarks, and digitisation. They include Sanskrit as one language among many.
Institutional digitisation projects — including a major collaboration in Chennai between IIT Madras and Madras Sanskrit College, working through more than 110,000 rare manuscripts to build a native Sanskrit model. Real and valuable, but weighted toward preservation and digitisation rather than a polished usable model.
BharatGen and other indigenous platforms — broad Indic coverage.
The smaller efforts¶
Paramanu-Sanskrit (around 139 million parameters) and similar small from-scratch models
ByT5-Sanskrit, focused on sandhi splitting and syntax parsing
Llama-family fine-tunes adapted for Sanskrit generation
Assorted cleaned Sanskrit datasets on public platforms
Why this is good news for you¶
They are doing the expensive part. Digitisation, benchmark creation, and corpus cleaning are slow, costly, and mostly published openly. You do not have to repeat any of it.
They are optimising for breadth. A model covering 22 languages cannot go deep on one. That is not a criticism, it is what their goal requires.
Nobody is doing depth on classical Sanskrit. Precise sandhi handling, faithful translation of shastra, metre-aware generation, philosophical analysis that does not fabricate — this is unclaimed.
Nobody is doing neuro-symbolic seriously. The Panini angle in Step 15 has very little competition, largely because it requires both engineering and domain knowledge, and few people have both.
The honest risks¶
Data is still the main limit. GRETIL, SARIT, and DCS are real resources, but cleaning, sandhi handling, and scholarly validation take genuine work.
You will need scholarly help for evaluation. In philosophical and religious material, a fluent wrong answer is worse than no answer. You cannot judge this alone unless you are already a Sanskritist.
Competition exists at the broad Indic level. If your plan is “a general Indic model”, you will lose. If your plan is “the best tool for splitting compounds in classical texts”, you can win.
Your model will be superseded. Plan for it. This is why Step 25 says your evaluation set and tokenizer may outlast your weights.