Enterprise AI firm M37Labs, which operates across India and San Francisco, has announced the launch of Saransh (सारांश), a small language model purpose-built to convert long-form Indian news content into concise summaries.
Founded in October 2024 by Prashant Shivram Iyer and Zorawar Purohit, M37Labs developed Saransh as the first model under its Enterprise Proprietary Model (EPM) methodology. Rather than fine-tuning an existing foundation model, Saransh was designed, tokenized, and trained from scratch using the company's own NVIDIA H100 infrastructure. Its architecture and weights were developed in-house using a corpus of Indian English news articles and reference summaries.
The model is specifically tailored to the linguistic and editorial characteristics of Indian news. Regional financial and institutional terms—such as crore, lakh, SEBI, MPC, and the GST Council—are processed as native domain vocabulary rather than adapted layers on top of a Western-trained general-purpose model.
Saransh has been designed around a single, narrow task: summarizing Indian news articles. It is not structured to act as a general-purpose chatbot, answer questions, write code, or execute arbitrary instructions. According to the company, this constraint is intentional, allowing the model to be evaluated, monitored, and governed against predictable behaviors in regulated and security-sensitive environments.
Co-founder and CEO Prashant Shivram Iyer stated that Saransh represents the practical application of building narrower, owned, and governed models deployable inside enterprises. Co-founder and Chief AI Officer Zorawar Purohit added that designing the model from the ground up for a specific task establishes a different risk and governance profile compared to general-purpose alternatives.
Saransh is the initial release in M37Labs' broader EPM development pipeline. The company is actively developing additional vertical small language models targeted at the Retail, BFSI, and Healthcare sectors, with releases planned over the next six months. The firm sees its EPM framework as a repeatable methodology to create domain-specific models optimized for enterprise workflows without data egress, ensuring predictable economics and controlled deployment.
Saransh is currently available for enterprise evaluation and can be deployed on-premises or within a client-controlled cloud environment. M37Labs is completing a published benchmark against a held-out evaluation set to form part of the model's technical documentation ahead of general availability.
"The launch of Saransh highlights a shift toward narrow, domain-specific artificial intelligence models tailored for regional and industry-specific requirements. By focusing on sovereign small language models trained from scratch for specific workflows, companies like M37Labs are addressing enterprise concerns around data governance, security, and predictability. As vertical models roll out across sectors like BFSI and healthcare over the coming months, enterprises will likely evaluate these controlled deployments as a viable alternative to general-purpose foundation models." — Dr. Shishir Gupta, Founder & CEO, StartupLanes
