Glint Research

We teach tiny neural networks to think. Sometimes they surprise us. Sometimes they just output zeros.

What We Do

• Design and train million-parameter generative models
• Study the behavior of small architectures under constrained compute

Models

Anthos — Tiny Flower Generation

Glint — 1M Parameter Model Series

Shard — 10M Parameter Model Series

More models are in active development. Follow us to stay updated!

Associates

We collaborate with other researchers and builders who share an interest in small models.

TeichAI

Armand0e

Crownelius

Enderchef

Principles

Transparency. Every model we release includes architecture details, training configurations, loss curves, and known limitations. We do not publish numbers we cannot reproduce.

Efficiency. We target few-GPU, short-run experiments. A model that trains in 8 hours on consumer hardware is more compact than one that requires a cluster.

Honesty about scale. Small models have hard limits. We document them clearly rather than overstating capability.

Community

Research is better with others. If you are working on small generative models, efficient training pipelines, or compact architectures and want to exchange ideas, we maintain an active Discord server.

Support

Glint Research is an independent, self-funded effort. If you find the work useful, support via Ko-fi helps cover compute costs and keeps experiments running.

Glint Research — We teach tiny neural networks to think. Sometimes they surprise us. Sometimes they just output zeros.