The models behind the products

Keepvidya Core — our own models.

Small models that run on modest hardware, built and tuned in-house. The same models power Flows, and they're how Keepvidya keeps your work yours — no third party in the loop.

Small by designFrom ~1M to 7B — sized to run, not to impress. Runs on-deviceModest CPUs — private by default, offline-friendly. Built in-houseTuned or trained from scratch by Keepvidya research.
Three families Shiva Shiva-Nano Ganesha
Language · powers Flows

Shiva

Our working language models — the ones writing your storybook prose and your game code today.

Tuned by Keepvidya research

Shiva-Chat

Ready

Conversation, summaries, and the storybook prose — the everyday assistant model.

Two sizes · 1.5B and 7B

Shiva-Code

Ready

Writes the small, self-contained games and snippets you generate in Flows.

Two sizes · 1.5B and 7B

Shiva-Writer, Novelize & Story

Ready

A trio of writing specialists — turning dry topics into explainers, books, and short stories.

Size · 1.5B
Language · from scratch

Shiva-Nano

Our smallest language models, written and trained from the ground up — proof we can build our own, not just tune someone else's.

Built from scratch, in-house

Shiva-Nano 1.0 & 1.1

Trained

The first ones we wrote by hand — tiny, complete, and trained end to end.

Size · about 1M parameters each

Shiva-Nano 2.0

Trained

A larger from-scratch run — the step up in size, trained all the way through.

Size · 27.8M parameters

Shiva-Nano 3.0

Not trained yet

The next size up — the architecture is ready; the training run hasn't happened yet.

Size · about 100M parameters
Vision & generative · from scratch

Ganesha

Our in-house work on images — models that understand pictures, and, in time, ones that paint them. A matured Ganesha will one day illustrate the storybooks Flows makes.

Built from scratch, in-house

Ganesha-Vision

Trained

Understands what's in a picture — an encoder built on JEPA (Joint-Embedding Predictive Architecture).

Trained · 73.87% accuracy on the CIFAR-10 benchmark

Ganesha-Gen

Not trained yet

Generates new images — a DDPM (Denoising Diffusion Probabilistic Model). One day it'll paint the storybooks; today it's built and waiting on its first training run.

Built · not trained yet

Ganesha-Text

Planned

Text understanding for search and retrieval — also a JEPA encoder. On the roadmap, not yet started.

Planned · no code yet

Every spec and status here is real — including what isn't trained yet. Honesty is a brand pillar, not a marketing line. Numbers update only when they're true.

See the models at work

Flows runs free on these models — paste any plain text and get back a finished storybook or a playable game.

Try Keepvidya Flows