Home · Models

The Model Estate

Named sovereign model families. Fine-tuned in our factory, deployed in your perimeter.

Open weights in, owned intelligence out. Five domain-tuned families — healthcare, banking, retail, pharma and Indic voice — built on permissively-licensed bases, quantised for your hardware, and run where your data already lives. Nothing rented by the token. Nothing leaves the building.

Run a family on your data →How sovereign AI works

Named families

Live & stable in production

4-bit

Quantised in-house

Tokens rented, ever

ChCharaka 1.5HealthcareLive & Stable LkLekha 2.5Banking, financial services & insuranceLive & Stable VnVanij 1.0Retail, consumer & supply chainLimited Production RsRasayanaPharma & life sciencesIn Development BhBhasha 1.0Indic voice & vernacularPOC Phase

How it works

Open weights in, owned intelligence out — governed at every layer

One pipeline produces every family. We start from open-weight bases we are licensed to own, tune and distil them on curated domain corpora, quantise them for the GPUs and edge devices our clients actually have, and ship them inside the accelerators that put them to work. Governance is enforced at each step, not audited afterwards.

Open-weight bases

Apache-2.0 / MIT-class model families — weights we are licensed to own, modify and redistribute inside a client perimeter.

Model factory

Tune · distil · quantize. Domain fine-tuning on curated corpora, distillation to the smallest size that holds accuracy, 4-bit quantisation for client GPUs and edge devices.

Canonical domain models

Charaka · Lekha · Vanij · Rasayana · Bhasha — the named families, each with a golden evaluation set and a provenance record.

117 accelerators

The production-tested workflows that put a model to work: onboarding, claims, clinical documentation, procurement, voice.

Inside your perimeter

Deployed on your hardware, your identity provider, your SIEM. Zero egress. Nothing rented by the token.

The families

Five domains. One factory. Named because they are products, not projects.

Each family carries its own golden evaluation set, a provenance record for every deployment, and a version that stays frozen in production until you decide otherwise.

Live & Stable

Charaka 1.5

Healthcare

Clinical documents · revenue cycle · specimen & diagnostics workflows

Explore Charaka →

Live & Stable

Lekha 2.5

Banking, financial services & insurance

Onboarding · KYC · fraud signals · claims straight-through processing

Explore Lekha →

Limited Production

Vanij 1.0

Retail, consumer & supply chain

Procurement · vendor & supply-chain documents · SAP + ServiceNow orchestration

Explore Vanij →

In Development

Rasayana

Pharma & life sciences

Regulatory submissions · clinical-trial documents · pharmacovigilance · quality SOPs

Explore Rasayana →

POC Phase

Bhasha 1.0

Indic voice & vernacular

Offline Indic voice Q&A · vernacular text · edge devices

Explore Bhasha →

Governance

Defensible in diligence, auditable in production

Provenance register

Base model, version and licence recorded for every deployment — the first artefact an auditor or a procurement team asks for.

Golden set per domain

Every family carries a curated evaluation set drawn from the document types its clients run on. Releases are gated on it.

Blind human + AI-judge scoring

Outputs scored by domain reviewers without knowing which model produced them, cross-checked by an AI judge. Raw runs available on request.

Frozen in production

The model that served your customers on Friday is the model serving them on Monday. No silent upgrades, no vendor deprecations, no pricing changes.

What a family is — and isn't. Families are product-line brands for domain fine-tunes, not claims to from-scratch pretraining. Named families ride Apache-2.0 / MIT-class bases; the base model, version and licence behind every deployment are recorded in a provenance register. The value we sell is the curated corpus, the evaluation discipline, the quantisation for your hardware, and the fact that it runs inside your perimeter.

FAQ

Questions buyers and auditors ask

What is the MindMap Model Estate?

A set of named, domain-tuned sovereign model families — Charaka (healthcare), Lekha (BFSI), Vanij (retail and supply chain), Rasayana (pharma) and Bhasha (Indic voice) — produced by one factory pipeline: permissively-licensed open-weight bases, fine-tuned and distilled on domain corpora, 4-bit quantised, and deployed inside the client's own perimeter. Two families are live and stable in production; the rest are in limited production, POC or development.

Are these models trained from scratch?

No, and we say so plainly. The families are product-line brands for domain fine-tunes of Apache-2.0 / MIT-class open-weight bases. The value is in the curated domain corpora, the evaluation sets, the quantisation for client hardware, and the provenance discipline — not in pretraining a foundation model. Base model, version and licence for every deployment are tracked in a provenance register.

How is a sovereign model different from calling a cloud API?

The model moves in with you. Weights sit on hardware you control, inference runs inside your network, logs land in your SIEM, and there is no per-token bill and no API to phone home — the stack can operate fully air-gapped. A cloud API is rented capability with your data doing the commuting; a sovereign model is owned capability with the data staying put.

Can I evaluate a family on my own documents before committing?

Yes — that is the standard first step. We run the relevant family against a sample of your real documents or transcripts on infrastructure you control, score it against your own acceptance criteria, and share the results and the evaluation harness. Request an evaluation from any family page.

Evaluate before you commit

Run a family on your own documents.

The standard first step, and the only honest one. Send us 20–50 representative documents or transcripts — anonymised is fine — and your own acceptance criteria. We run the relevant family on infrastructure you control and hand back the scored results and the evaluation harness itself.

Two weeks, fixed scope, no production data
Your GPUs or a sealed environment we provision — never a public API
You keep the harness whether or not we proceed