AI research from first principles. By @jiteshkprasad
Joined July 2026
- Tweets53
- Following1
- Followers568
- Likes64
Lekhani now covers Telugu.
Type in English letters and it writes Telugu script on your own device, from a download under 1 MB.
Lekhani by @udgamlabs now writes Telugu. Type "meeru ela unnaru" in English letters and get మీరు ఎలా ఉన్నారు, right in your browser.
81% of words right on the first guess, in under 1 MB. That's 8.5 points ahead of IndicXlit's model.
1
6
170
Whisper large-v3 wrote words on 71% of clips where nobody speaks. Three other open models wrote nothing on any of them.
We played 950 clips to 27 speech-to-text settings, most of them silence, noise, music, coughs and laughs, and counted what they wrote. The most common invention: "Thank you." The same Whisper turbo weights wrote words on 58%, 83% or 87% of no-voice clips depending on the library.
The fix is cheap: with a voice detector in front, every model fell to 1.8% or less. Every transcript is public.
Write-up: udgamlabs.com/blog/hush-benc…
Board: huggingface.co/spaces/udgaml…
Shruti now writes all 22 languages of India's Eighth Schedule, each from one file under 30 MB, with English in the same file.
Anyone can use it on their own device with no sign-in. Developers can integrate it with no key, no account and no server, and all 22 are also on ONNX, through pip and npm, at about 32 MB and 54 MB.
The code and the technical report are already out. Built on Cactus Whistle. Open weights, Apache-2.0.
Happy Diwali. Shruti by @udgamlabs now types all 22 languages of India's Eighth Schedule with no internet, each in a file under 30 MB, about a fifth of IndicConformer's 137.7 MB.
It is an AI model that writes text from speech, and it runs on a five-year-old phone. If you speak one of these 22, you can use it on your own phone, right in the browser.
If you build apps, you can integrate it, and I want any developer to be able to serve millions of Indians in their own language without depending on a big company, a cloud bill or a big GPU. Open weights, Apache-2.0, built on Cactus Whistle.
Shruti now writes Sanskrit, Punjabi, Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri, Nepali, Santali and Sindhi, in five scripts, each from one 29.7 MB file, offline on a phone or laptop.
On a five-year-old phone: about 9x to 13x faster than real time, depending on the language, in 64 MB of memory (native engine, 20 clips each). Open weights, Apache-2.0.
Shruti by @udgamlabs adds 11 languages today. It is an AI model that writes text from speech, and each language is one 29.7 MB file that runs on your own device.
7 of the 11 aren't on OpenAI Whisper's language list: Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri and Santali. The other 4 are Sanskrit, Punjabi, Nepali and Sindhi.
5
402
Shruti Assamese types Assamese and English from one 29.7 MB file, offline on a phone, laptop or browser.
On a five-year-old phone: about 11x faster than real time, in 64 MB of memory (native engine, 20 clips). Open weights, Apache-2.0.
Shruti Assamese by @udgamlabs is out. It is an AI model that writes text from speech, Assamese and English, in one 29.7 MB file that runs on your device.
On Assamese conversations that mix in English, it makes about 24% fewer errors than IndicConformer-600M, a model about 86 times its size. That is 30.2 against 39.8 word errors per 100 words on 84 turns, with English words counted only in English letters.
1
2
240
Shruti Malayalam types Malayalam and English from one 29.7 MB file, offline on a phone, laptop or browser.
On a five-year-old phone: about 10x faster than real time, in 64 MB of memory (native engine, 20 clips). Open weights, Apache-2.0.
6 English words in one Malayalam sentence, and all 6 stayed in English letters on my five-year-old phone: "ഞാൻ airport ലേക്ക് ഒരു cab book ചെയ്തു driver location share ചെയ്തിട്ടുണ്ട്". That means I booked a cab to the airport, the driver has shared the location.
Shruti Malayalam by @udgamlabs is live. It is an AI model that writes text from speech, Malayalam and English, in one 29.7 MB file that runs on your device.
Shruti Urdu writes Urdu and English on one line, offline on a phone, laptop or browser. About two thirds of the English words stay in English letters.
On Urdu conversations that mix in English: word error rate 22.9, IndicConformer Urdu 29.1, on 74 turns. Open weights, Apache-2.0.
We released Shruti Urdu at @udgamlabs, an AI model that writes text from speech. On Urdu conversations with English mixed in, its word error rate is 22.9 and IndicConformer Urdu's is 29.1, on 74 turns.
Urdu goes right to left and English goes left to right, and people use both in one sentence, right? So about two thirds of the English words stay in English letters, on the same line.
Kannada and English in one 29.7 MB file. On Kannada conversations that mix in English, Shruti Kannada makes fewer errors than IndicConformer Kannada, a model almost five times its size, and Whisper Kannada fine-tunes up to 33 times its size.
Offline on a phone, laptop or browser. Open weights, Apache-2.0.
29.7 MB against 137.7 MB, and on Kannada conversations with English mixed in, the smaller one makes fewer mistakes. That smaller one is Shruti Kannada by @udgamlabs, an AI model that writes text from speech, and the other is IndicConformer Kannada.
I think this is how people actually talk: "ಮನೆ owner ಹತ್ತು ತಿಂಗಳ ಬಾಡಿಗೆ advance ಕೇಳ್ತಿದ್ದಾರೆ". In the video that line is the real output on my phone, word for word.
Shruti Odia is now live: Odia and English speech to text in one 29.7 MB file, offline on a phone, laptop or browser.
On real Odia conversations that mix in English, counting English words only in English letters, fewer errors than every model we measured. Open weights, Apache-2.0.
36.2% word errors, against 43.4 for a model about 31 times its size. Shruti Odia by @udgamlabs is now live, and on real Odia conversations with English mixed in, it makes fewer mistakes than every other open model we measured.
It is an AI model that writes text from speech, Odia and English, in one 29.7 MB file, offline on your phone, laptop or browser.
1
1
172
PhoneGPT is training live right now, on a 2021 phone's own chip. No cloud.
It has about 11 hours to go below 1.4697. The score updates on the stream after every checkpoint.
A $330 phone from 2021 is trying to go below a data-centre GPU's score, live, with @udgamlabs.
It's called PhoneGPT, a nanoGPT-style model with about nine times fewer parameters, and it runs on the phone's own chip with no cloud.
The score to go below is 1.4697 from nanoGPT, which took an A100 about 3 minutes. This phone gets about 11 hours.
As far as we can find, nobody has done this live on a phone; correct us.
teapawt.net/i/broadcasts/2XyELyJLB…
Shruti Gujarati is live: Gujarati and English speech to text in one 29.7 MB file, offline on a phone, laptop or browser. English words stay in English.
On real conversations that mix in English, fewer errors than every model we measured. Open weights, Apache-2.0.
Shruti Gujarati by @udgamlabs is live. One 29.7 MB AI model, and on real Gujarati conversations with English mixed in, it makes fewer mistakes than every model we measured, including one 30 times its size.
It writes Gujarati and English, offline, on your phone, laptop or browser. The English words stay in English, the way people actually speak.
4
352
Shruti Telugu is live: Telugu and English speech to text in one 29.7 MB file, offline on a phone, laptop or browser. English words stay in English.
On real conversations that mix in English, fewer errors than Telugu models 5× and 30× its size. Open weights, Apache-2.0.
Shruti Telugu is 29.7 MB AI model by @udgamlabs. On real Telugu conversations with English mixed in, it makes fewer mistakes than Telugu models five and thirty times its size on speech to text.
One file that writes Telugu and English, offline, on your phone, laptop or browser. The English words stay in English, the way people actually speak Telugu.
Shruti Bengali is out: Bengali and English speech to text in one 29.7 MB file, offline on a phone, laptop or browser. English words stay in English.
On real conversations that mix in English, fewer errors than Bengali models 3× and 5× its size. Open weights, Apache-2.0.
Shruti Bengali by @udgamlabs is live. One 29.7 MB small AI model that writes Bengali and English text from speech, offline, on your phone, laptop or browser.
We all mix English into Bengali. It writes those English words in English, and on real conversations like that it makes fewer mistakes than Bengali models three and five times its size.
Shruti Tamil: Tamil and English speech to text in one 29.7 MB file, running offline on your phone, laptop or browser. English words stay in English.
The most accurate open Tamil speech-to-text model under 100 MB that we measured. Open weights, Apache-2.0.
Shruti Marathi: Marathi and English speech to text in one 29.7 MB file, running offline on the phone. English words stay in English.
The most accurate open Marathi speech-to-text model under 100 MB that we measured. Open weights, Apache-2.0.
1
4
439
9 AI models in 31 days. This was October at Udgam Labs.
udgamlabs.com/models
Can a useful AI model fit in under 1 MB?
Lekhani by Udgam Labs: Hindi typing in 0.87 MB, in your browser. 65.9% on the Aksharantar Hindi test, 7 points over IndicXlit's model; its full package is still ahead (68.7%).
udgamlabs.com/models/lekhani
3
12
620
Shruti writes Hindi, English and Hinglish speech from one 29.5 MB file, offline, on your own device.
Try it in your browser: udgamlabs.com/models/shruti
Ten days ago @udgamlabs released Shruti, a 29.5 MB speech-to-text model for Hindi, English and Hinglish. It's 100x smaller than OpenAI's Whisper large-v3, more accurate on Hindi, and runs offline in your browser.
And people are already building with it, which is exciting to see.
4
426
Drishya Tiny by Udgam Labs is out: a 212M video generator trained from scratch on one GPU for about $21 of GPU time, using the frozen Wan2.2 VAE and Flan-T5-Base text encoder. Give it a text prompt, or a photo plus a prompt, and it makes a 2-second 256×256 clip.
On VBench picture quality it scores 75.6, level with VideoCrafter-0.9 (74.9) and close to ModelScope (78.1). It follows prompts about half as well (32.3 vs 65-67), and we say so up front. VBench here is our own sampling and scoring with the official code.
This is a research release: weights, code, logs, the full cost ledger ($57 with every failed run) and what didn’t work are all open.
Run it in your browser, on your own device: udgamlabs.com/models/drishya…
Technical report: udgamlabs.com/research/drish…
Drishya Tiny is out. It's a 212M video generator we trained from scratch, on one GPU, for about $21 of GPU time. Give it a text prompt, or a photo plus a prompt, and it makes a 2-second clip.
On VBench its picture quality is level with older video models built on Stable Diffusion, one of them with a network about six times its size. Following the prompt is a different story, more on that below.
The weights, the code and every failed run are public. And you can run it right in your browser, on your own device, nothing gets sent anywhere.
3
357
Vemana by Udgam Labs is out: a 278M Telugu base model, trained from scratch, live, on one GPU.
0.361 bits per byte on 416 Telugu Wikipedia pages from Sep–Oct 2026, the lowest of every open base model under 1B we tested.
Run it in your browser: udgamlabs.com/models/vemana
Vemana is out. It's a 278M Telugu model we trained from scratch, live, on one GPU, for about $69.
And I think the result surprised even me: on new Telugu Wikipedia pages it beats every open base model under 1B that we tested, including nanochat-telugu-560M, which is twice its size.
The weights, every checkpoint and the eval code are public. And you can run it right in your browser, on your own device, nothing gets sent anywhere.
8
1,023