AI & ML interests
AI research from first principles.
Recent Activity
Articles
Paar by Udgam Labs: 12 small open translation models, English into 12 languages, 17.7 MB each, offline on a CPU or in the browser. Apache-2.0.
Lekhani: type 12 languages in Latin letters and get their own script, under 1 MB each, in your browser, on your own device.
Vemana: a 278M Telugu base model trained from scratch, live in public, for about $69. Weights, checkpoints, GGUF and demo.
Tests and open data for small language models. Scoring code: github.com/udgamlabs/udgam-evals
Solves English maths word problems by writing a short Python program and running it. Trained from scratch; runs on a laptop CPU.
English on-device models for tool calling, trained from scratch. 384M parameters; run in Ollama, LM Studio, llama.cpp and MLX.
Offline speech to text, ~30 MB each, with English, in 22 languages: one small model per language. Phone, laptop, browser.
212M video generator trained from scratch (frozen Wan2.2 VAE, Flan-T5-Base). 2-second 256×256 clips from text, or text plus a first frame.
Laghu: a 152M chat model trained from scratch for $146. Weights, checkpoints, data, evals and demo.
384M English model trained from scratch on ~17B tokens. Thinks step by step, holds a conversation, reads ~16K tokens. Apache-2.0.
Text-to-image models trained from scratch by UdgamLabs. Chitra (चित्र) means "picture".
Our first releases, kept public for transparency, including negative results.
Paar by Udgam Labs: 12 small open translation models, English into 12 languages, 17.7 MB each, offline on a CPU or in the browser. Apache-2.0.
Offline speech to text, ~30 MB each, with English, in 22 languages: one small model per language. Phone, laptop, browser.
Lekhani: type 12 languages in Latin letters and get their own script, under 1 MB each, in your browser, on your own device.
212M video generator trained from scratch (frozen Wan2.2 VAE, Flan-T5-Base). 2-second 256×256 clips from text, or text plus a first frame.
Vemana: a 278M Telugu base model trained from scratch, live in public, for about $69. Weights, checkpoints, GGUF and demo.
Laghu: a 152M chat model trained from scratch for $146. Weights, checkpoints, data, evals and demo.
Tests and open data for small language models. Scoring code: github.com/udgamlabs/udgam-evals
384M English model trained from scratch on ~17B tokens. Thinks step by step, holds a conversation, reads ~16K tokens. Apache-2.0.
Solves English maths word problems by writing a short Python program and running it. Trained from scratch; runs on a laptop CPU.
Text-to-image models trained from scratch by UdgamLabs. Chitra (चित्र) means "picture".
English on-device models for tool calling, trained from scratch. 384M parameters; run in Ollama, LM Studio, llama.cpp and MLX.
Our first releases, kept public for transparency, including negative results.