One Family, Three Footprints
There is no single right size for a sovereign AI model. A model that fits comfortably on a workstation in a branch office is the wrong choice for a national data centre, and a cluster-scale model is useless inside a vehicle with one GPU and no internet.
That is why PRAGYA, the Tosh.AI model family, is not a single model. It is three tiers - Edge, Core, and Max - that share the same lineage, the same Indic and domain depth, and the same sovereign principles, but target very different deployment footprints. You pick the tier that matches your hardware and your latency needs, not the other way around.
All three are built on fine-tuned open-weight foundations. That choice is deliberate. Open weights mean the model is inspectable, runs on your own hardware, and carries no dependency on a foreign endpoint. We then fine-tune for Indic-language fluency and domain knowledge so the model performs where generic foreign models are thin.
PRAGYA Edge
Edge is the tier for places where compute is scarce and connectivity cannot be assumed. It is designed to run on a single GPU, a capable laptop, or compute mounted inside a vehicle, and it runs fully offline.
This is the tier for the field. A team operating in a remote location, a device in a moving vehicle, or a workstation inside an air-gapped facility can run real AI locally, with no round trip to a server and no network dependency at all. Edge trades raw scale for portability and resilience, and for many frontline tasks that is exactly the right trade.
Because it is fully offline, Edge also offers the strongest privacy guarantee in the family. Nothing leaves the device.
PRAGYA Core
Core is the on-premise workhorse, and for most organisations it is the default. It is sized to run inside a customer's own data centre or server room on a sensible amount of hardware, serving a department or an entire organisation.
Core is where the everyday work happens: drafting and summarising, answering questions grounded in internal documents, powering internal assistants, and serving as the reasoning engine behind agentic workflows. It balances capability and cost so that a single on-premise deployment can support real production load without reaching for a cluster.
When Core is paired with our grounding layer, it answers from your own documents with citations, and when paired with YANTRA it becomes the engine behind multi-step tasks. It is the tier most customers build their platform around.
PRAGYA Max
Max is the top tier, built for cluster deployment and the most demanding work. It targets large context windows and heavier reasoning, so it can take in long documents, large case files, or extensive retrieved context and reason across all of it at once.
Max is for organisations with serious infrastructure and serious problems: complex analysis over large corpora, high-volume workloads, and tasks where the extra capability and context length materially change what is possible. It runs across a cluster of GPUs in the customer's own environment, keeping the same sovereign, air-gap-capable posture as the smaller tiers - just at much greater scale.
What They Share
The tiers differ in size, but they share what makes PRAGYA sovereign. Every tier is built on open-weight foundations you can inspect. Every tier runs inside your own perimeter with zero foreign dependency. Every tier is air-gap capable. And every tier carries the Indic-language and domain fine-tuning that lets it perform on Indian content and specialised subject matter where generic models fall short.
This shared lineage matters in practice. You can prototype on Core, push a distilled workload to Edge in the field, and reserve Max for the heaviest central analysis - all with consistent behaviour and a single trust model across the family.
Choosing a Tier
The selection rule is straightforward. Start from the hardware and the constraints. If you need fully offline operation on minimal compute, Edge fits. If you are deploying on-premise for a department or organisation, Core is the workhorse. If you have cluster infrastructure and large-context, high-volume needs, Max is built for it.
You can see how PRAGYA fits within the wider platform - alongside Sovereign RAG and YANTRA - on our models page, and explore deployment options on our enterprise page.
To talk through which tier suits your environment, contact us.