Agentic Foundation
Daisy Nova
Production-grade agentic foundation model. Powers planning, tool-use and multi-step workflows across every domain agent.
- Parameters
- Mixture-of-Experts
- Context
- 1M tokens
- Languages
- EN · HI · MR
- Residency
- Maharashtra
Sovereignty · Onshore by design
Onshore weights. Onshore keys. Onshore audit.
Maharashtra-resident compute, state-held HSM keys and an append-only ledger — CAG-ready, RTI-disclosable, reversible.
G2G · Mantralaya Copilot
Mantralaya clears a 240-page inter-ministry file in one sitting.
The copilot summarises every annexure, surfaces precedent and flags policy conflicts — with full CAG-ready lineage.
G2C · Aarogya Agent
An ASHA worker in Gadchiroli gets a second opinion in 30 seconds.
Aarogya Agent screens symptoms against PMJAY protocols and books a teleconsult when the PHC is out of reach.
G2C · Vidya Tutor
Every Zilla Parishad student gets a Marathi tutor that never tires.
Vidya Tutor explains a math step ten different ways — and tells the teacher exactly where each child is stuck.
Foundation models
Frontier models, Marathi-first, Maharashtra-resident.
A family of foundation models tuned for Indian context, Indian languages and Indian regulation — trained, fine-tuned and served on sovereign Maharashtra infrastructure.
FAMILY
10 frontier models
One MoE core. Nine specialists. All onshore.
LANGUAGES
मराठी · हिंदी · English
+ 18 Marathi dialects, 12 Indic scripts.
RESIDENCY
maharashtra-r1
Weights, embeddings & logs never leave the state.
PROVENANCE
Daisy Nova lineage
Proven in commercial Insure AI in North America.
What is a frontier model
A frontier model is the most capable system of its generation — built at the limits of compute, data and method.
Scale
Hundreds of billions of parameters trained on trillions of tokens. Enough capacity to hold law, language, code, science and Marathi dialects in a single model.
Generality
One base model that transfers to thousands of downstream tasks — reasoning, retrieval, planning, vision, voice — without retraining from scratch.
Emergence
Capabilities that only appear past a certain scale: tool-use, multi-step reasoning, chain-of-thought, cross-lingual transfer, in-context learning.
Foundation · Frontier · Specialist
Three tiers in the MahaAI family — each with a job.
TIER 01
Foundation models
Daisy Nova · MahaAI Marathi Frontier
Trained from scratch on raw, multilingual, multimodal corpora. These are the base weights every other system inherits from. Expensive to build, cheap to reuse.
TIER 02
Frontier capabilities
Voice · Vision · Code · Compliance
Mid-training and post-training on the foundation base to unlock a specific modality or skill at state-of-the-art quality — streaming speech, document OCR, code synthesis, legal reasoning.
TIER 03
Domain specialists
Krishi · Aarogya · Bhumi · Lokshahi
Small, distilled, instruction-tuned models — fine-tuned on departmental knowledge graphs. Cheap to run on the edge, governed by the same audit and residency rules.
Anatomy of a MahaAI foundation model
Six parts. Each one engineered, audited and resident in Maharashtra.
01
Tokenizer
Devanagari-aware byte-pair tokenizer with 256K vocabulary. Marathi compounds, Sanskrit roots, code and emoji share one address space — no script penalty for Indic users.
02
Architecture
Decoder-only transformer with Mixture-of-Experts routing on Daisy Nova; 70B dense on Marathi Frontier. RoPE positions, grouped-query attention, sliding-window cache for 1M context.
03
Pre-training corpus
4.2T Marathi tokens, 3.8T Hindi, 9T English, 1.1T code, 600B legal and gazette text. Provenance-tagged, deduplicated, PII-scrubbed, licence-cleared.
04
Alignment
SFT on 1.8M Marathi instruction pairs curated with SCERT and IIT Bombay. DPO plus constitutional rules drawn from the Indian Constitution, DPDP Act and ministry SOPs.
05
Safety & evals
Bharat-Bench, IndicGenBench, MMLU-MR and an internal 'Mantralaya Hard' eval. Red-teamed against caste, communal, electoral and procurement-fraud attack surfaces.
06
Serving
Quantised to FP8/INT4 on Maharashtra-resident GPU clusters. Speculative decoding, paged KV cache, per-tenant isolation. Every inference logged to the sovereign audit ledger.
Training pipeline
From raw Marathi text to a deployable agent in five stages.
01
Corpus
Curate, dedupe, licence-clear, PII-scrub. Provenance hashed for every shard.
02
Pre-train
Trillion-token runs on sovereign GPU mesh. Checkpoints signed, weights never exported.
03
Mid-train
Long-context, multimodal and tool-use extensions on top of the base.
04
Align
SFT, DPO, constitutional AI with Indian legal and cultural rule packs.
05
Serve
Quantise, route, log. Every token traceable to a policy and an officer.
The model family
Six models in production. Each tuned for a job the state actually does.
Language
MahaAI Marathi Frontier
First Marathi-native frontier LLM — trained on 4.2T tokens of Marathi corpus, vetted with SCERT and IIT Bombay linguists.
- Parameters
- 70B dense
- Context
- 256K tokens
- Languages
- Marathi-first, 11 Indic
- Residency
- Maharashtra
Speech
MahaAI Voice
Sub-300 ms voice agent stack for call-centres, IVR and field-officer apps. Robust to rural dialect variation.
- Parameters
- Streaming ASR + TTS
- Context
- Realtime
- Languages
- MR · HI · EN + 18 dialects
- Residency
- Maharashtra
Multimodal
MahaAI Vision
Document understanding for 7/12 extracts, claim forms, satellite crop imagery and CCTV review.
- Parameters
- Vision-language
- Context
- Image + 128K
- Languages
- Multilingual OCR
- Residency
- Maharashtra
Code
MahaAI Code
Code assistant fine-tuned for DigiLocker, India-Stack APIs, and the state's legacy COBOL/RPG estate.
- Parameters
- 34B dense
- Context
- 200K tokens
- Languages
- Polyglot
- Residency
- Maharashtra
Specialist
MahaAI Compliance
Rule-pack model that turns gazette notifications, GRs and SOPs into machine-checkable compliance graphs.
- Parameters
- Distilled 13B
- Context
- 512K tokens
- Languages
- Legal EN/MR
- Residency
- Maharashtra
Benchmarks
State-of-the-art on Indic — competitive with global frontier on English.
92.4
IndicGenBench-MR
best-in-class Marathi generation
88.1
MMLU (EN)
general knowledge, 5-shot
94.7
Bhasha-Voice
ASR accuracy, rural Marathi
96.2
Bhumi-OCR
7/12 extract field accuracy
For ministries & departments
Bring sovereign AI to your department.
Pilot a Daisy-powered agent inside your ministry in 90 days, with audit-grade decision logs and full Maharashtra data residency.