Tokz Labs is now a member of the NVIDIA Inception program.
We build small open models that make an AI agent's routine decisions: which team gets a ticket, whether a message is urgent, which tags apply. You send text and a typed question, and you get back a probability for every option without generating any text.
Our models run on NVIDIA GPUs end to end. OpenDecision was trained on an H100, and the Tokz API serves every request on NVIDIA L4s.
What the program gives us
Inception members get NVIDIA developer resources and training, exclusive pricing on NVIDIA hardware and software, and exposure to the venture capital community. For a lab our size, the compute pricing and the training resources matter most.
What comes next
- A bigger model. We are fine-tuning a 31B model on NVIDIA H200s and will submit it to the Jev Decision Index. We will publish the score once it is official.
- Label sets we have not seen. On label sets that no training data targeted, the 4B models in our technical report still generalize better than OpenDecision. Closing that gap is the main goal of the next model.
- The same open approach. OpenDecision weights stay on Hugging Face under Apache-2.0, and the API stays pay per token.
If your agent makes routine decisions, try the API. New accounts start with $5 of free credit.