Our models.
Built in the open.
A new current for AI. A human-led community building toward models we train ourselves—and tools everyone can run.
More people should
have a hand in
what comes next.
Building AI should be something people can participate in, understand, and take with them.
Tidebots brings the work into one place: source files, datasets, model cards, checkpoints, and the conversations behind them. Our ambition is a community that builds its own models and helps people run them locally.
We are starting small. The collaboration foundation comes first. A tiny, reproducible training experiment comes next.
Where we are today
Formation stage. No trained model released. No hosted training service. The roadmap is our proposed path.
A home for the model.
And everything behind it.
Software collaboration and model stewardship, in the same conversation.
↳ README.md
↳ model-card.md
↳ data/provenance.json
↳ training/config.json
↳ evaluation/results.json
First Current
A small language model, trained from random weights on reviewed human-written material.
Create your own project →Independence starts
with the foundations.
Know what teaches it.
Begin with a small, reviewed collection of human-written material. Record its source, permissions, transformations and hash. Exclude known model-generated training text.
No OpenAI or Anthropic base models, teachers, or known generated training outputs.A small beginning.
A direction worth building.
Form the foundation
Define the idea, build the workspace, invite the first contributors.
Train the first baseline
Review a small corpus. Train from random weights. Measure validation results.
Make it reproducible
Release code, source records, results and a checkpoint with clear reuse terms.
Grow the commons
Improve models together. Expand provenance tools, local runners and review.
The next current
starts with people.
A small team. Clear first tasks. Work other people can check and build on.
Start a project ↗Creates a draft for you to share. Nothing is submitted or sent.
Open questions.
Clear intentions.
Can I use a Tidebots model today?+
Not yet. Our proposed first model is a tiny from-scratch experiment to prove the pipeline. We will show actual results when they exist.
What can the workspace do now?+
Create private or community-visible projects, edit source files with revision history, maintain cards, discuss work, and store small checkpoints. Full Git hosting, pull requests, team editing and hosted training are future work.
What does independent mean here?+
The intended model pipeline excludes OpenAI and Anthropic base models, teachers and known generated training outputs. The hosted website preview uses ChatGPT sign-in for access; model training and inference do not use their APIs.
How will you check the data?+
Record sources, permissions, transformations and hashes. Public corpora may contain undisclosed synthetic text. Review reduces this risk without establishing a perfect guarantee.
How does licensing work?+
The website source uses MIT. Model weights and datasets need separate permissions and explicit licenses. Download access alone does not grant reuse rights.