PRAGYA MODEL FAMILY

Sovereign models, built defence-first.

PRAGYA is the reasoning and intelligence model family that powers Tosh.AI - the best available open-weight foundations with continued pretraining and fine-tuning on Indic and domain data. The family is designed around deployment realities rather than benchmark vanity: foundation models in three sizes, plus specialised variants for defence, government, enterprise, and multilingual work.

Foundation models

Three sizes, one family

From a disconnected workstation at a forward base to a sovereign cluster - pick the size that fits the deployment, not the other way round.

PRAGYA 7B

Laptops · Field systems · Edge

Lightweight and quantised. Runs on a single GPU, a strong CPU, or a laptop - built for forward-deployed, air-gapped, fully offline operation where connectivity cannot be assumed.

PRAGYA 13B

On-prem GPU box

General-purpose enterprise and government reasoning. The standard sovereign-deployment workhorse for grounded retrieval and everyday agentic workflows inside your own perimeter.

PRAGYA 30B

On-prem cluster · Sovereign cloud

Advanced reasoning and large-context workloads. For complex analysis and multi-step YANTRA workflows at scale on a cluster or single-tenant sovereign cloud.

More than a model - a platform

The model is one component. The orchestration is the product. PRAGYA ships with grounding and agentic execution, not just a chat box.

Sovereign RAG Grounding

PRAGYA answers from your own documents - classified within your own perimeter - with citations. Access-controlled retrieval over your vector store, fully offline.

YANTRA Agentic Orchestration

A planner decomposes a goal into steps, a tool router connects PRAGYA to internal systems through allow-listed tools only, and specialist agents coordinate complex, multi-step tasks reliably.

Indic + Defence-Domain Depth

Continued pretraining and supervised fine-tuning on Indic-language corpora, defence and government doctrine, technical manuals, and tool-use trajectories.

See PRAGYA running inside your perimeter.

From first conversation to a running, air-gapped demo - in 30 days.

Request Demo