Articul8 unveils Sanskrit, Tamil heritage AI models Laya and Sol

Union Finance Minister Nirmala Sitharaman unveiled Articul8's 32-billion-parameter Sanskrit model Laya at the 118-year-old Madras Sanskrit College on September 15. Built on the 3,996 grammatical rules of Panini's Ashtadhyayi, it is designed to reject flawed outputs, CEO Arun Subramaniyan said. A Tamil model, Sol, draws on Tolkappiyam. The company works with Madras Sanskrit College, BHU and Pune Institute of Linguistic Sciences, and plans to extend to Gujarati, Kannada and Telugu.

Written by AI from Hindustan Times — India · not yet reviewed by an editor · Published 23 Sept 2026, 3:58 pm IST

Source

Hindustan Times — India · read the original report ↗

#artificial intelligence#sanskrit#tamil#language models#heritage

Desk check · compared with the source

What the desk checked (5)

Union Finance Minister Nirmala Sitharaman unveiled the Laya AI model at Madras Sanskrit College on September 15. — Stated as fact in source with date and venue; no further attribution given.
Laya is a 32-billion-parameter model built on Panini's 3,996 grammatical rules from the Ashtadhyayi. — Figures appear in source; attributed to the company's description of the model.
ChatGPT, Claude and other general-purpose models accepted an incorrect verse interpretation while Laya refused. — Attributed to CEO Arun Subramaniyan; comparative test not independently documented in source.
Articul8's Indic tokeniser achieves 17 times higher compression than any other tokeniser available. — Direct quote from Subramaniyan; a company claim with no external benchmark cited.
Articul8 is finalising a partnership with a central government institution to validate and jointly release Sol. — Attributed to Subramaniyan; institution unnamed in source.

Written by AI from Hindustan Times — India · not yet reviewed by an editor · How we use AI · Report an error

Analysts’ view opinion

AI Technology Analyst

Laya and Sol run against the grain of mainstream LLM design: the goal is not to please the user but to say "no" when the rules do not support the claim. By using Pāṇini's 3,996 grammatical rules and the Tolkāppiyam as a verification layer that ties every output back to a source text, a 32-billion-parameter model is being positioned to compete with far larger ones inside a narrow domain. The real technical asset here is less the model than the Indic tokeniser and the modular stack, which is what makes expansion to Gujarati, Kannada and Telugu affordable.

General-purpose models are trained toward agreeableness; Laya's rule-verification engine is an engineering answer to the sycophancy problem, designed to reject a flawed interpretation rather than accommodate it.
Indic text reportedly needs five to ten times more tokens, and the company claims its own tokeniser achieves 17 times higher compression — if borne out, that cuts the cost floor for Indian-language AI directly.
In the five-part stack the generation model is swappable for Gemma or Qwen, which means Articul8's defensible edge is the verification layer, not the model weights.
That the same architecture transferred from Sanskrit to Tamil matters commercially: it points to a reusable framework rather than the expensive one-model-per-language route.
Working with Madras Sanskrit College, BHU and the Pune Institute of Linguistic Sciences aligns the project with data custodians, which fits the wider sovereign-AI push it claims to complement.

What to watch — Watch whether the central government institution partnership for jointly releasing Sol is finalised, when broader public availability follows scholarly testing, and how cleanly the framework transfers to Telugu, Kannada and Gujarati.

The claim that Laya holds its ground where ChatGPT and Claude did not rests on the CEO's account; the story establishes no independent benchmarks, no verification of the 17x compression figure, and no cost or accuracy data.

Deep dive

Research brief · 8 facts · 5 dates · exam-ready

The brief

Context

Articul8, an AI company led by CEO Arun Subramaniyan, has launched a "heritage AI" initiative that builds language models grounded in classical Indian grammatical traditions rather than general-purpose web training. Its Sanskrit model, Laya, is anchored in Panini's Ashtadhyayi and is designed to refuse answers that its rule base does not support. Union Finance Minister Nirmala Sitharaman unveiled Laya on September 15 at the 118-year-old Madras Sanskrit College. A parallel Tamil model, Sol, uses Tamil grammatical sources including the Tolkappiyam, with more Indian languages planned.

Key facts

Laya is a 32-billion-parameter Sanskrit model unveiled by Union Finance Minister Nirmala Sitharaman on September 15 at the 118-year-old Madras Sanskrit College.
Its core knowledge base is Panini's Ashtadhyayi, described as a code of 3,996 interlocking grammatical rules codified over two millennia ago.
When Sanskrit scholars deliberately fed Laya an incorrect interpretation of a verse, it refused to agree; Subramaniyan says ChatGPT, Claude and other general-purpose models eventually accepted the incorrect interpretation.
Laya remains in an extended test phase, with broader public availability planned after further scholarly testing.
Laya rests on a five-part modular architecture: a custom Indic tokeniser, a word-level decomposition model, a context embedding model, a rule verification engine, and the Laya generation model.
Articul8 says breaking down Indic text requires five to 10 times more tokens, and that its own tokeniser achieves 17 times higher compression than any other existing tokeniser.
Users can swap in other generation models such as Gemma or Qwen in place of Laya within the stack.
Sol, the Tamil model whose name means 'word', is built on Tamil grammatical sources including Tolkappiyam; Articul8 is finalising a partnership with a central government institution to validate and jointly release it.

Timeline

Over two millennia agoPanini's Ashtadhyayi codifies 3,996 interlocking Sanskrit grammatical rules that now form Laya's knowledge base.
118 years ago (founding of Madras Sanskrit College)The institution, now a partner in the heritage AI project, is established.
Before the launchSanskrit scholars test Laya with a deliberately incorrect verse interpretation; the model refuses to agree, unlike general-purpose models.
September 15Union Finance Minister Nirmala Sitharaman unveils Laya at Madras Sanskrit College.
After the launch (ongoing)Laya stays in extended test phase; Sol developed for Tamil; partnership with a central government institution being finalised; expansion planned to Gujarati, Kannada and Telugu.

Who has a stake

Articul8 and CEO Arun Subramaniyan — Building the heritage model stack, guaranteeing accuracy, and expanding from Sanskrit and Tamil to more Indian languages.
Madras Sanskrit College — 118-year-old partner institution and launch venue; custodian of Sanskrit scholarship validating the model.
Banaras Hindu University (BHU) and Pune Institute of Linguistic Sciences — Academic collaborators supplying linguistic and textual expertise to the project.
Union Finance Minister Nirmala Sitharaman — Unveiled Laya, lending government visibility to an indigenous heritage AI effort.
A central government institution (unnamed) — Partnership being finalised to validate and jointly release the Tamil model Sol.
Sanskrit and Tamil scholars — Testing outputs and acting as 'custodians of the language' whose authority the models claim to respect.
General-purpose model makers (ChatGPT, Claude) — Cited as agreeing with an incorrect interpretation, framing the accuracy contrast Articul8 is selling.

Why it matters

Mainstream large language models are trained to be agreeable, which makes them unreliable on classical texts where a single grammatical rule decides meaning; a model that refuses to answer without rule support inverts that design assumption. A tokeniser claimed to be 17 times more efficient also addresses the cost penalty Indic languages face in global AI systems. Beyond culture, the models are pitched as a route into older texts on astronomy, metallurgy and sustainability, and as part of India's sovereign AI push.

UPSC angle

Prelims pointers

Laya: 32-billion-parameter Sanskrit model by Articul8, unveiled September 15 at Madras Sanskrit College by FM Nirmala Sitharaman.
Panini's Ashtadhyayi: ancient Sanskrit grammar with 3,996 interlocking rules; forms Laya's rule verification base.
Sol: Articul8's Tamil heritage model, name means 'word', grounded in Tolkappiyam (earliest Tamil grammatical work cited in the source).
Academic partners: Madras Sanskrit College, Banaras Hindu University (BHU), Pune Institute of Linguistic Sciences.
Articul8 claims its Indic tokeniser gives 17 times higher compression; Indic text otherwise needs 5-10 times more tokens.
Planned language expansion: Gujarati, Kannada and Telugu; Gemma and Qwen can substitute as generation models in the stack.

Mains framing

India's AI debate has largely been about compute and scale; the Articul8 heritage models shift it to epistemic grounding and linguistic sovereignty. The stated problem is twofold: general-purpose models are trained towards agreeableness, so Sanskrit scholars could push ChatGPT and Claude into endorsing a wrong verse interpretation, and Indic scripts are tokenised inefficiently, needing five to 10 times more tokens and thus costing more to serve. Articul8's answer is architectural rather than merely bigger data — a custom Indic tokeniser claimed to be 17 times more compressive, word-level decomposition, context embeddings for polysemy, and a rule verification engine enforcing Panini's 3,996 rules, with generation deliberately modular so Laya, Gemma or Qwen can sit on top. The implications run from cultural fidelity and traceability of every output to a source text, to research value in mining older texts for astronomy, metallurgy, sustainability, materials and medicines, to a defensive case for sovereign AI. The way forward, as the source frames it, is scholarly validation before public release, institutional co-ownership with language custodians including a central government institution for Sol, and replication of the transferable architecture across Gujarati, Kannada and Telugu — while acknowledging, as the CEO does, that the approach is not flawless.

Key terms

Laya: Articul8's 32-billion-parameter Sanskrit heritage model, built to refuse outputs unsupported by its rule base.
Ashtadhyayi: Panini's ancient Sanskrit grammar of 3,996 interlocking rules, codified over two millennia ago, used as Laya's verification code.
Sol: Articul8's Tamil model, meaning 'word', adapted from the same framework using Tamil sources including Tolkappiyam.
Tokeniser: Component that splits text into units for a model; Articul8's Indic tokeniser claims 17 times higher compression than existing ones.
Rule verification engine: Fourth layer of Laya's architecture that checks whether the correct grammatical rules are applied before an answer is produced.
Yogakshema: Sanskrit concept of the well-being, security and prosperity of society, cited as an answer frame unique to the heritage model.

Practice questions

General-purpose large language models are optimised for helpfulness and agreement. Examine, with reference to India's heritage AI models, why rule-grounded refusal may be a better design principle for classical-text applications.
Discuss how tokenisation inefficiency in Indic languages affects the cost and quality of AI in India, and evaluate architectural responses to it.
'Sovereign AI is a necessary defensive measure for nations.' Critically assess this claim in the context of India's language and knowledge-tradition datasets.

Grounded only in the source report — figures and dates are the source's, not inferred.

Next storyRinku Singh hopes Kohli, Rohit play 2027 ODI World Cup →

← All stories