01 · The line
AI research from first principles
First principles means we start from why something works, not from what is popular. Before we build, we work out what the data, the architecture and the training are each doing, whether we are training a model, fine-tuning one or applying AI to a new problem. Then we publish what we measured, including where we fall short.
02 · How we think
Three principles, each with an example from our work
- 01
Understand before we build
We work out why something works (the data, the architecture, the training) before we scale it, whether we are training a model, fine-tuning one or applying AI to a new problem.
From our workSmall models make arithmetic slips when they reason in text, so Lilavati writes a short Python program and lets Python do the arithmetic. At 95M parameters it solves 54.7% of the GSM8K test on its first try. Lilavati
- 02
Honest numbers
Every result names its benchmark, its settings and what it is compared with, and we publish where we are behind.
From our workGargi comes within 2 points of Qwen2.5-0.5B on ARC-Easy, ARC-Challenge and WinoGrande, and the same table shows it near chance on MMLU: 25.5 against Qwen2.5-0.5B’s 47.3. Gargi’s results
- 03
Small and efficient by design
We build models that run on a laptop or a phone and cost a fraction of the usual amount to train.
From our workGargi learned from about 17 billion tokens, about 1,000× less than Qwen2.5-0.5B’s ~18 trillion. Chitra Tiny cost $100.46 of GPU compute for the whole project, trial runs included. Chitra Tiny’s compute ledger
03 · The name
Where the name comes from
Udgam (उद्गम) is Sanskrit for origin: the source of a river. The mark draws that source as one seed and two streams that fork once more: one, two, four.
India is where the lab began, and our models carry names from its history: Gargi after the Vedic philosopher Gargi Vachaknavi, and Lilavati after the book of arithmetic problems that the mathematician Bhāskara II wrote in 1150 CE.
04 · Next
Where we are going
Today the work is research: small models, each with a model page and a technical report that states every number and every limitation.
Next come products, custom models for businesses, and AI in the physical world. The products we already run keep their own names and are listed on the Products page.
05 · Who we are
Who we are
Udgam Labs was founded in 2024 by Jitesh Kumar Prasad, who leads its research and products. The lab operates through two companies: Udgam Labs LLP in New Delhi, India, where the work began, and Udgam Labs, Inc., its US subsidiary, incorporated in Delaware.
- 6
- Models released
- 4
- Technical reports
- 2024
- Founded