Efficient by design.
Treepai is building specialized language models, each meant to do a specific task very well. Our goal is models that are cheap to train, cheap to run, and built on data we can account for.
of training on one consumer GPU, from scratch. Our second model, Treepai g1, averages 49.6% on five core benchmarks, ahead of GPT-2 medium (44.9%) and Pythia-410M (47.0%), the established open models of its size.
against a Pythia-410M checkpoint at a similar 10.5 billion training tokens. g1 leads on every core benchmark.
Fine-tuned to sort customer-support messages, g1 gets 95.2% of categories right, against 95.6% for Qwen2.5-0.5B fine-tuned the same way.
It is not the best model of its size. Qwen2.5-0.5B and SmolLM2-360M, trained with hundreds of times more data, are ahead on general benchmarks, and g1 was the slowest of four models in our speed test.
Read the reportMost work doesn’t need the biggest model.
Sending every request to a large, general model is expensive for work that’s specific and repetitive. A model built for one task can be much smaller, and much cheaper to run. Our goal is specialized models that work alongside frontier AI, not against it.
Our first research result is consistent with this. In a controlled test at 110M parameters, replacing part of the model’s educational training text with broad web text raised everyday common sense and lowered science, rather than improving both. It is one test at one size, and we will repeat it on larger models. See the report.
Built for one job
We’re building toward models that each handle a specific task, from small specialists to large task models.
Runs where you need it
Our goal is models that can run on your own servers, including ordinary CPUs with no GPU, or through a hosted API.
Efficiency first
Cost to train and cost to run are design goals from day one, not something we fix later.
Honest evaluation
We won’t claim a result until we can show it and explain how we measured it. For g0, the benchmarks and the rule for what counts as a win were fixed before the runs. How we measured.
Clean data
Trained from scratch on documented data with clear provenance, including AI-generated examples that carry an open license and whose origin we can verify.
Treepai is a member of the NVIDIA Inception program.
NVIDIA and the NVIDIA logo are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.