What does Treepai do?
Treepai is an AI lab and model provider. We research and train specialized language models, each meant to do a specific task very well, and to cost less to train and run than large general models. More about us.
Is anything available to use?
Not yet. Our first model, g0, is a research baseline. Our next model, g1, is in training.
Are g0’s weights released?
No. g0’s weights, tokenizer and training code are not released. Its model card says what it is and isn’t suitable for.
How good is g0?
On five core benchmarks it averages 37.3%, against 38.6% for GPT-2 small, and 38.9% with our improved training recipe. SmolLM2-135M, trained on far more data, is well ahead. The numbers, the method and the limits are in the report.
How did you measure it?
With one evaluation harness for every model, benchmarks fixed before training, and a rule for what counts as a win set in advance. How we measured.
Can my company work with you?
We’re building toward models for specific business tasks. If you have high-volume, repetitive language work and want to talk, send us a message.
How do I follow your work?
Enter your email in “Research updates” at the bottom of any page, or subscribe to the Updates feed.
How do I cite the report?
The citation and a BibTeX entry are on the report page.
What is RealUptime?
RealUptime is a company Treepai acquired in October 2026. It continues as its own company with its own products. The announcement.
Are you hiring?
Not yet. Open roles will be posted on the Careers page first.
How is the name pronounced?
“Tree-pie”. Where it comes from.
I’m writing about Treepai. Where are the logos?
On the Press and brand page, with a press contact.