BharatGen

Patram: vision language models for India

Patram is our family of vision language models. It looks at an image, whether a scanned form, a photographed bill, a chart or a street sign, and understands it the way a person would.

  • 7 billion parameters
  • Forms, tables, charts and stamps
  • Open weights
Why Patram

Why Patram: India's paper trail, understood

India runs on documents: forms, orders, certificates, bills and records, in many scripts and formats. Much of it is still read by hand.

India runs on paper

Government orders, land records, insurance policies, bank statements, medical reports and court filings. Most arrive scanned or photographed, with stamps, signatures and handwriting.

Reading is not understanding

OCR turns a page into text and leaves a second system to make sense of it. Patram reads the image and answers directly: which dates changed, what the table says.

Built on Indian formats

Trained on real Indian documents, from Aadhaar formats, IFSC codes and cheques to stamps, seals and government forms, as they actually arrive: scanned or photographed.

Documents stay private

Documents often carry personal data that should not leave the building. Patram's weights are open, so it can run inside your own systems, on your own servers.

Patram‑7B‑Instruct releases

Patram 1, 1.1 and 1.2 are revisions of one 7 billion parameter model on Hugging Face. Each release carries forward everything before it.

June 2025

Patram 1

Our first vision language model. Answers questions about a page directly, with no separate OCR step, and handles summaries and code-mixed documents.

Built for
Aadhaar formats, IFSC codes, cheques, stamps and seals, government forms
August 2025

Patram 1.1

Adds multi-turn conversations about a document and key-value extraction.

Adds
Multi-turn Q&A, key-value extraction
February 2026

Patram 1.2

Adds step-by-step reasoning, text extraction, and understanding of charts, tables and flowcharts.

Adds
Reasoning, OCR, charts and tables

Patram 1 to 1.2 are English-first, with some ability on Indian scripts.

How it reads

Built to read the way humans do

Patram takes in the whole image at once: headings, tables, stamps and signatures, and how they sit together. Then it answers in words.

Illustration of how a vision language model reads a page.

One model, no OCR step

From pixels to words, in one pass

The image is cut into small patches. Patram's vision encoder turns each patch into something its language model can read, and the answer comes out word by word.

Illustration. Each square is one patch of the image.

Patram 1.2 on Indian documents

Two real answers: reasoning through an insurance claim form, and reading a scanned government notice word for word.

Image and question
A blank Janata Personal Accident Insurance claim form from United India Insurance, with instructions and four numbered sections
Open the full-size image ↗

"Tell me how to fill it, what all fields are necessary and how many max claims can i make?"

Patram 1.2 answers

Thinking step by step:

  1. The document contains a section titled 'Important Instructions' which outlines the procedure for filling the form.
  2. The instructions state that the form must be filled out completely and accurately, with 'Not Applicable' or 'N/A' written where questions do not apply.
  3. The instructions also specify that the treating doctor's certificate or medical report (Annexure 1) must be submitted.
  4. The form includes a section 'IV. NATURE OF CLAIM' with checkboxes for different types of claims, and a note stating that only one of the listed benefits can be claimed for the same injury.
  5. The form requires details in sections I through IV, including claim intimation, policyholder details, insured person details, and nature of claim.

Fill out the form completely and accurately, write 'Not Applicable' or 'N/A' where questions do not apply, submit the treating doctor's certificate or medical report (Annexure 1), and select only one nature of claim from the options provided.

Insurance claim form · Step-by-step reasoning

Real output, not edited. Source: Patram 1.2 model card ↗

Build with Patram

The weights are open: download Patram from Hugging Face and run it on your own servers, or talk to us about reading your documents with it.