Alexandria-based startup TokenAI has launched Neo, a compact AI decision model designed to choose the right tool for a task or escalate the decision to a human reviewer.
Neo has 41 million parameters and was trained from scratch on 400,000 synthetic records covering about one million decisions. It can produce three outputs in a single pass: a choice from up to 32 options, a score across up to five levels, or a binary yes/no decision.
The model is built to identify low-confidence decisions instead of executing a tool automatically. TokenAI describes this feature as “confidence-aware routing,” aimed at helping AI agent pipelines handle routing and escalation logic quickly and with limited computing resources.
On a held-out benchmark, Neo recorded 73.99% accuracy on its score output and 82.38% on binary decisions. Its choice accuracy across 32 options was 36%. Running on CUDA hardware, it can process about 10,241 decisions per second, with latency below 0.1 milliseconds.
TokenAI has released Neo’s weights, training code, dataset and documentation on GitHub and Hugging Face. The TokenAI Neo Model License prohibits commercial, client-facing or production use without written permission, keeping the release focused on research for now.
Source: Middle East AI News
