M37Labs, an enterprise AI company operating across India and San Francisco, announced the availability of Saransh (सारांश), a small language model purpose-built to turn long-form Indian news content into concise, accurate summaries.
Saransh is the first model developed under M37Labs’ Enterprise Proprietary Model (EPM) methodology and marks the beginning of a planned family of vertical small language models. The company is currently developing dedicated models for Retail, BFSI and Healthcare, with releases targeted over the next six months.
Unlike a fine-tuned version of an existing foundation model, Saransh was designed, tokenized and trained from scratch, with its architecture and weights developed in-house. Training was conducted on M37Labs’ own NVIDIA H100 infrastructure using a corpus of Indian English news articles and reference summaries.
The model was designed specifically around the linguistic and editorial characteristics of Indian news. Terms such as crore and lakh, as well as references to institutions including SEBI, MPC and the GST Council, are learned as native domain vocabulary rather than being treated as adaptations layered onto a general-purpose Western-trained model.
The EPM (Enterprise Proprietary Model) Thesis
Saransh is more than a summarization model. It is the first working demonstration of a broader proposition from M37Labs: enterprises can create greater long-term business value through AI by owning smaller, purpose-built models trained for specific workflows and data and governance deployed entirely within their own technology perimeter.
M37Labs calls this category the Enterprise Proprietary Model, or EPM.
The EPM approach is built around three core principles:
- Predictable economics: A small, purpose-built model running on dedicated infrastructure can transform variable, metered inference costs into a more predictable infrastructure cost at enterprise scale.
- Data sovereignty: Models deployed within an enterprise’s own network or cloud environment can process sensitive documents and proprietary information without requiring data to leave the organization’s perimeter.
- Reproducibility and control: Frozen, versioned model weights allow enterprises to reproduce outputs and maintain a consistent model version over time, rather than depending on continuously changing hosted models.
“Saransh is the first model coming off an AI Native firm like M37labs. We are building narrower, owned, governed models deployable inside the enterprise. That is the M37Labs EPM thesis we are helping Enterprises build for their long term AI transformation.”
— Prashant Shivram Iyer, Co-Founder & CEO, M37Labs
“A model designed for one task has a fundamentally different risk and governance profile from a general-purpose model being asked to perform that task. We chose not to fine-tune our way to Indian news fluency. Saransh learned the language, vocabulary and patterns of its domain from the beginning of training. That is a harder engineering path, but we believe it is the direction enterprise AI is heading.”
— Zorawar Purohit, Co-Founder & Chief AI Officer, M37Labs
What Saransh Does — and Doesn’t
Saransh has been deliberately designed around a single task: summarizing Indian news articles.
It is not intended to be a general-purpose chatbot or reasoning model. It does not answer questions, write code, conduct conversations or execute arbitrary instructions. Its narrow scope is intentional.
M37Labs views this constraint as an advantage for enterprise deployment. A model optimized for a clearly defined task can be evaluated, monitored and governed against a specific set of expected behaviors — an important consideration for regulated and security-sensitive environments.
Potential applications include:
- Media monitoring at scale: Converting daily coverage across hundreds of publications into concise, searchable digests without accumulating a per-document inference charge.
- Leadership and communications intelligence: Compressing days or weeks of media coverage into executive-ready briefs.
- Newsroom and wire support: Generating story digests and standfirsts within a publication’s own infrastructure.
- Archive processing: Summarizing years of historical content in bulk — particularly valuable where metered, per-token inference economics become prohibitive.
M37Labs ‘Saransh’ at a Glance
|
Attribute |
Specification |
|
Class |
Small Language Model |
|
Primary Task |
Abstractive summarization |
|
Domain |
Indian news & media |
|
Language |
Indian English |
|
Model Lineage |
Trained from scratch |
|
Architecture & Tokenizer |
Developed in-house |
|
Training Hardware |
NVIDIA H100 |
|
Inference Footprint |
Single GPU |
|
Deployment |
On-premises or client cloud |
|
Data Egress |
None |
|
Governance |
Frozen, versioned weights + model card |
|
Status |
Available for evaluation |
One Model Factory. Four Verticals.
Saransh is the first model from M37Labs’ EPM development pipeline, not the last.
The company has confirmed active development of additional vertical small language models for Retail, BFSI and Healthcare, with releases targeted over the next six months.
|
Vertical |
Initial Focus |
Strategic Foundation |
|
Retail |
Catalogue, category and consumer-signal summarization; merchandising and product-content compression |
Builds on M37Labs’ RetailIO.AI deployment experience |
|
BFSI |
Regulatory filings, research notes, persistency information and advisory correspondence |
Extends M37Labs’ BFSI and insurance/wealth AI workstreams |
|
Healthcare |
Clinical documentation and patient-communication summarization; India-context medical language |
New vertical for M37Labs |
The company sees the EPM factory as the larger business opportunity.
Rather than building one large model and attempting to make it relevant to every enterprise workflow, M37Labs is developing a repeatable methodology for creating small, domain-specific models optimized for clearly defined enterprise tasks.
Across verticals, the company expects the same fundamental advantages to apply: predictable economics at scale, no data egress, controlled deployment and reproducible model behavior.
These characteristics become particularly important in environments such as BFSI regulatory documentation, healthcare records and communications, and high-volume retail catalogues and consumer signals.
Availability
Saransh is available now for enterprise evaluation and can be deployed on-premises or within a client-controlled cloud environment.
M37Labs is currently completing a published benchmark against a held-out evaluation set. The benchmark will form part of the model’s technical documentation ahead of broader general availability.
As with M37Labs’ EPM approach, model weights are frozen and versioned for each production release, with accompanying governance and model documentation.
