M37Labs launches Saransh as first sovereign small language model in planned India-focused family

Saransh and the Enterprise Proprietary Model pitch

M37Labs announced Saransh (सारांश), a small language model designed, tokenised and trained from scratch on the company's own infrastructure rather than fine-tuned from an existing foundation model. The company said Saransh is purpose-built to turn long-form Indian news content into concise summaries and is the first release under its Enterprise Proprietary Model (EPM) methodology, which argues that enterprises derive more durable value from narrower, owned models deployed inside their own perimeter than from general-purpose hosted systems.

How the model was built

Training was carried out on M37Labs' own NVIDIA H100 infrastructure using a corpus of Indian English news articles and reference summaries, with architecture and weights developed in-house. The company said vocabulary such as crore, lakh and references to institutions including SEBI, the MPC and the GST Council were learned as native domain terms during initial training rather than layered on later. CEO Prashant Shivram Iyer framed Saransh as the first model from what he called an "AI Native" firm, while co-founder and chief AI officer Zorawar Purohit said the company chose the harder engineering path of training from first principles over fine-tuning to reach Indian news fluency.

Deliberate limits on capability

Saransh is scoped to a single task, summarising Indian news, and is not intended to operate as a general-purpose chatbot, question-answering system, coding assistant or conversational agent. M37Labs said the narrow scope is deliberate: a model optimised for one defined task can be evaluated, monitored and governed against a specific expected behaviour set, which the company argues matters most in regulated and security-sensitive enterprise environments.

EPM principles and the vertical roadmap

The EPM framework is built on three stated principles: predictable economics from converting variable inference costs into steadier infrastructure spend, data sovereignty by keeping processing inside an enterprise's own network or cloud, and reproducibility through frozen, versioned weights. M37Labs said it is developing dedicated vertical models for Retail, BFSI and Healthcare, with releases targeted over the next six months.

A second sovereign AI launch on the same day

Separately on 18 August, ShepHertz Technologies announced AgentAnywhere, a sovereign agentic AI platform that runs inside customer infrastructure and is backed by seven model families trained in India, each named in Sanskrit. The first family, Taksha for AI engineering and coding, is generally available, with Manthan, Kuber, Seva, Tatva, Astra and Sanjaya slated through 2026 across Fast, Pro and Max tiers, and a parallel Indic-language line called Manthan Vaani in development.

What remains to be seen

The near-term milestones are the planned Retail, BFSI and Healthcare Saransh-family releases from M37Labs within the next six months, alongside the staged rollout of additional ShepHertz model families through 2026. Both companies have positioned sovereignty, in-house training and customer-side deployment as the defining design choices, leaving benchmark performance and enterprise uptake as the measures that will test those claims.