AI research from first principles.
Udgam Labs is an independent AI research lab. Before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Every result we publish comes with honest numbers.
New · Shruti · speech to text · 19 October 2026
100× smaller than OpenAI’s Whisper large-v3, and better at Hindi. It runs in your browser.
Shruti writes down Hindi, English and Hinglish from one 29.5 MB file. On the same Hindi test clips it makes fewer word errors than Whisper large-v3 (Monsoon 20.5 vs 33.3, FLEURS 20.1 vs 28.9, Lahaja 22.9 vs 31.2), a 3,087 MB model. On English, Whisper is clearly better. Press Load once: the model downloads to this browser and everything after that happens on your device, even with the internet off.
Every result, including where it is behind Demo on Hugging Face Code and examples on GitHub
Shruti · checking this device
Audio never leaves your device. Shruti and its open-source engine run inside this page.
(or drop one on the transcript)
Load Shruti, then press Record and speak Hindi, English or both.
The language is detected for you. English words stay in English (Latin script) and are underlined here; Hindi is written in Devanagari. While you speak, words appear once they have settled, a second or two behind you; when you stop, Shruti reads the whole recording once more for the final text.
Research index · 10 models
Small models. Measured progress.
Text to video · 28 Oct 2026
Drishya Tiny
Makes 2-second 256×256 clips from a sentence. A research release: about $21 of GPU time, every failure included.
- 212M generator
- 15 h on one L40S
- CC BY-NC 4.0 weights
Model page
Telugu language model · 27 Oct 2026
Vemana
A Telugu base model trained live in public. On fresh Telugu text, the best open base model under 1B that we tested.
- 278M parameters
- About 16 h on one H100
- Apache-2.0
Model page
Chat model · 22 Oct 2026
Laghu
A small English chat model with a thinking switch, trained from scratch for $146 and released with everything open.
- 152M parameters
- $146 pretraining run
- Apache-2.0
Model page
Speech to text · 19 Oct 2026
Shruti
Writes down Hindi, English and Hinglish, offline, on a phone or in the browser. English words stay in English.
- 29.5 MB
- Runs in the browser
- Apache-2.0
Model page
Language model · 17 Oct 2026
Gargi
A small English language model that thinks step by step on maths problems and runs comfortably on a laptop.
- 384M parameters
- ≈17B training tokens
- Apache-2.0
Model page
Maths word problems · 16 Oct 2026
Lilavati
Solves maths word problems by writing a short Python program and running it.
- 95M parameters
- GSM8K 54.7% first try
- Apache-2.0
Model page
Text to image · 12 Oct 2026
Chitra Tiny
Makes 256×256 pictures from an English sentence, on a laptop in about three seconds per image.
- 212M parameters
- 256×256
- Apache-2.0
Model page
On-device tool calling · 10 Oct 2026
Udgam-384M
Turns requests into function calls, documents into JSON and tasks into shell commands, on the device itself.
- 384M parameters
- Context 4,096 tokens
- Apache-2.0
Model page
Research release · 29 Sept 2026
Udgam English 125M
A small English model trained from random weights. It missed its performance target, and we published the full results.
- 125M parameters
- Context 2,048 tokens
- Target missed
Model page
Fine-tuned adapter · 22 Sept 2026
Udgam Hinglish Commands
Turns short English and Romanized Hinglish requests into a structured intent and details an app can use.
- LoRA on Qwen3-0.6B
- 54.88% vs 9.90% base
- 8 command domains
Walkthrough
Latest research
Reports and papers.
Technical report ·
Drishya Tiny: building a small video generator from scratch, failures included
Technical report for Drishya Tiny, a 212M-parameter video generator trained from scratch with the frozen Wan2.2 VAE and Flan-T5-Base: the design, the data, five stages with what each measured and what failed, VBench, a controlled test of what moves prompt-following, every dollar, and where it falls short.
Release postTechnical report ·
Vemana: a 278M Telugu language model trained from scratch
Technical report for Vemana, a 278M-parameter Telugu base model trained from scratch: data, tokenizer, model, the training run, a pre-registered test on new Telugu Wikipedia pages against every open base model under 1B we found, where it falls short, and the release.
Release postTechnical report ·
Laghu: a 152M chat model trained from scratch for $146, released fully open
Technical report for Laghu, a 152M-parameter English chat model trained from scratch for $146 of pretraining compute: architecture, data, the full training recipe, a pre-registered evaluation against size-class rivals, what did not work, the cost and the limitations.
Release postTechnical report ·
Shruti: offline speech-to-text for Hindi, English and Hinglish in 29.5 MB
Technical report for Shruti, a 29.5 MB speech-to-text model for Hindi, English and Hinglish built on Cactus Whistle: the Devanagari vocabulary extension, a full fine-tune with 4-bit quantisation-aware training, openly licensed training data with Hinglish labels, one evaluation harness for every model, measurements on a phone and in the browser, and limitations.
Release post
Three principles behind every release.
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.
Honest numbers
Every result names its benchmark, its settings and what it is compared with, and we publish where we are behind.
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.
Research collaborations and partnerships.
We welcome research collaborations, partnerships and questions about our models. Email is best for anything detailed. Replies usually within 24–48 hours.