title: RuVI Labs
emoji: 🚀
colorFrom: gray
colorTo: gray
sdk: static
pinned: false
RuVI Labs
Build good models for people.
A research lab from India · language models, built the honest way
01 · Who we are
RuVI Labs is a small, independent research lab based in India. We design, train and release language models end to end: architecture, data, training and inference, with no borrowed foundations.
We build for the people who will actually use the models: students, developers, and teams working with limited compute. If a model can't run where its users are, it isn't finished.
Small lab, serious work.
02 · Pragya
| Model | Status | Notes |
|---|---|---|
| Pragya Brahmã (flagship) | planned | Open weights, open data |
| Pragya Vayu | 🧪 Training | Built for low-resource hardware |
| Pragya medha | 🧪 Planned | Medium tier |
03 · Our principles
|
◼ Open weights Free to use, free to build on. |
◻ Open data Nothing hidden in the corpus. |
Capable release Release whats peak |
What "from scratch" means to us
- Own architecture. Designed and implemented in-house.
- Own training run. Pretrained by us on data we can document.
- Own tooling. Training, evaluation and inference code built and released alongside the weights.
What "open" means to us
- Weights you can download, run and modify.
- A data recipe you can inspect and question.
- Honest reporting, including what didn't work.
04 · What we believe
Good models are different.
I. Trained to perform differently. II. Built with the motive to undo what we assume is possible in AI. III. Made to achieve, and made to perform.
05 · What we're working on
- 🔨 Pretraining the Pragya family on open data
- 🧪 Efficient architectures that make small models punch above their size
- 🧪 Quantization and fast inference so models run on modest hardware
- 🧪 Open evaluation: transparent benchmarks and model cards
07 · Quick facts
| Lab | RuVI Labs |
| Founder | Arush Kumar |
| Home | 🇮🇳 India |
| Family | Pragya |
| License | Open weights |
| Mode | Building quietly. Shipping publicly. |
08 · FAQ
Are the models really trained from scratch? Yes. Architecture, pretraining and tooling are our own. Some Models are LoRA
Can I use the weights commercially? Check the license on each model card. Our goal is permissive, open release.
Can I contribute? Yes. Open a discussion on any repo here, or reach out below.
09 · Follow along
Follow the organization on Hugging Face to get notified when models and datasets drop.
RUVI LABS
small lab, serious work.