From the lab
Releases, research and notes from Udgam Labs — what we built, how we measured it and what we learned, in our own words.
Release ·
Chitra Tiny, a 212M-parameter text-to-image model trained from scratch
A small diffusion transformer that makes 256×256 pictures from a sentence, trained on about 4.7 million public-domain and CC0 images. Open weights, measured results and known failures.
Release ·
Udgam-384M, a small model for tool calling inside apps
A 384M-parameter English model trained from scratch to turn requests into function calls, JSON and shell commands on the device. Best at knowing when to act; behind on exact calls.
Release ·
Udgam English 125M, a small model that missed its target
We trained a 125.8M-parameter English model from random weights for about US$17 of cloud compute. It missed its performance target, and we released it anyway with every number.
Release ·
Hinglish Commands, a small parser for mixed Hindi–English requests
A LoRA adapter for Qwen3-0.6B that turns English and Romanized Hinglish commands into structured intents. It lifts exact match on Hinglish-TOP from 9.90% to 54.88%.
Follow new posts with the RSS feed.