Shunya AI
Agentic AI for India: fast, multilingual, enterprise‑ready.
Hosted in India • Zero data retention by default
Trusted by forward‑looking teams across BFSI, retail, telecom, manufacturing, and the public sector.
India's agentic AI—made for real work.
Shunya Mini 1.0 blends an enterprise‑grade language model with an integrated router and expert variants to deliver fast answers on simple tasks and deep reasoning when it matters.
What is Shunya Mini 1.0?
An agentic language model built for Indian users and organizations. It combines a dense Transformer core with specialized expert variants and an integrated router that chooses the right expert and tools for each step of a task—retrieval, reasoning, code/SQL, or structured JSON output.
Large context
Up to 128K tokens (best performance in 32K range) for long documents and multi‑file workflows.
Indic languages by design
Eighth Schedule languages + widely used non‑scheduled varieties; robust code‑switched Indian English and mixed‑script input.
Enterprise‑ready
Indian datacenters, zero data retention by default, serverless API only (no VPC/on‑prem), domain‑restricted RAG, and schema‑validated function calling.
Why it matters (for India)
Talk like India talks
Understands Hinglish and regional code‑switching across scripts.
Aligned with Indian policy
Engineered for DPDP Act, 2023 compliance support and MeitY due‑diligence advisories.
Operational at Indian scale
Optimized to serve high‑volume customer support, policy/notice summarization, and multi‑language field workflows.
Interoperability with India Stack
DigiLocker (read‑only, scoped)
Fetch user‑authorized documents for summarization and structured extraction (e.g., licences, certificates) without storing originals.
UPI rails via partners
Trigger status checks and payment‑adjacent workflows through partner integrations.
Account Aggregator (AA)
Generate plain‑language explainers for consent flows; extract structured signals from bank/NBFC statements where legally permitted.
ONDC‑style schemas
Map catalogs, normalize attributes, and generate bilingual product copy aligned to ONDC models.
All integrations require customer authorization and, where applicable, partner agreements.
What makes Shunya different
1) Agentic routing, not one‑size‑fits‑all
Shunya's router infers the task type, sets a thinking budget (Light / Medium / Complex), selects tools (RAG, function calls), and picks the right expert variant at each step—plan → execute → reflect.
2) Speed and depth
Speculative decoding and KV‑chaining give you snappy responses on routine queries, while the model slows down deliberately for critical reasoning and structured outputs.
3) Multilingual safety
Safety alignment covers Indian languages and code‑switch, with ML‑based detection to reduce harmful or disallowed content across scripts.
4) Built for governed enterprises
Per‑tenant logging/retention controls, automatic masking for Indian IDs (Aadhaar/PAN) by default, and serverless API delivery (no VPC/on‑prem).
Enterprise outcomes & benefits
Customer experience
- • Reduce average handle time with multilingual copilots and self‑serve FAQs.
- • Increase first‑contact resolution with grounded answers and structured outputs.
Operations & productivity
- • Automate form‑filling, reconciliation, and MIS queries (text‑to‑SQL).
- • Turn policy circulars and SOPs into concise checklists and templates.
Risk, compliance & governance
- • Default masking for Indian IDs (Aadhaar/PAN); schema‑validated tool calls.
- • Per‑tenant logging and auditability; zero retention by default.
How it works (at a glance)
Understand
The router analyzes your prompt, org context, policies, and available tools.
Plan & assign
It chooses a reasoning depth and dispatches the best expert (multilingual, retrieval, code/SQL, or formatting).
Execute & verify
Experts call tools, generate outputs, and self‑check with plan–execute–reflect loops.
Deliver & cite
For grounded tasks, the retrieval expert formats citations and reduces hallucinations.
Depth profiles:
Minimal deliberation for speed.
Balanced reasoning for everyday work.
Deep, multi‑step reasoning for high‑stakes tasks.
Architecture & training highlights
Model Architecture
- • Dense Transformer with SiLU activation, ~80 layers for deep contextual reasoning
- • Vocabulary ~155K tokens; context window up to 128K (optimal ~32K)
- • Model class: up to 40B‑parameter family (exact parameter count not disclosed)
- • Distilled variants available
Training Scale
- • Training scale ~18.5T tokens total
- • ~4T code (90+ languages), 3T+ Indic, 11T+ English/global
- • Sequence lengths: pretrained on 4K/8K, extended to 128K via RoPE
- • 50M+ instruction‑tuning samples; GRPO (stage 1) and DPO (stage 2) alignment
Frequently Asked Questions
Is Shunya open source?
No. Shunya Mini 1.0 is closed‑source and available via a serverless API only (no VPC/on‑prem).
Where is my data processed and stored?
Requests are processed in Indian datacenters with zero data retention by default. Optional logging/audit may be enabled by contract.
Does Shunya train on my data?
Enterprise deployments can disable training on customer prompts. See your DPA and contract terms.
Which languages are supported?
Eighth Schedule languages and widely used non‑scheduled varieties, plus Hinglish and mixed‑script input.
Can Shunya make API calls to our systems?
Yes—via function calling with JSON‑Schema contracts. You control scopes and guardrails.
How do you reduce hallucinations?
A retrieval expert optimizes grounding and citation formatting; schema‑validated tool calls constrain outputs.
What about pricing, SLAs, and rate limits?
Usage‑based pricing and enterprise SLAs are available—contact sales for details.
Ready to get started?
Join forward‑looking teams who are transforming their operations with India's agentic AI.