#Egypt #AImodels - Alexandria-based startup TokenAI has launched Neo, an open-weights System One Decision Model built to route tasks to the right tool, or escalate them for human review. The 41 million parameter decision model is trained from scratch on 400,000 synthetic records covering a million independent decisions. Neo scores calibrated choice, score and yes/no outputs in a single pass, and is designed to flag low-confidence decisions for review rather than act on them. The new model has been released with its weights, training code and dataset publicly available, under a custom TokenAI licence (that restricts commercial use).
SO WHAT? - Neo has been built to make one narrow decision reliably, which could be what tool to use, or whether a human needs to step in. So, this is developed to help for anyone trying to run routing logic cheaply and fast inside an AI agent pipeline. The restriction against production or commercial use keeps it firmly in research territory for now, but its release poses an interesting question: how small can a dedicated decision model be while staying calibrated?
TokenAI has released Neo, an open-weights System One Decision Model designed to select actions from a declared set of candidate options, rather than generate conversational text.
Neo has 41 million trainable parameters, built on a bidirectional Transformer encoder with 8 layers, a hidden size of 512 and 8 attention heads, trained entirely from scratch.
The model was trained on 400,000 records covering roughly one million independent decisions, generated by TokenAI’s internal synthetic data tool neo-english-synthetic-v1, and split into training, calibration and test sets.
Neo produces three types of output in a single pass: a choice among up to 32 candidate options; a score across up to 5 ordered levels; and a binary yes/no decision (useful for tasks like tool routing and escalation detection).
On a ‘held-out benchmark’, Neo scored 73.99% accuracy on its score head and 82.38% on its binary decision head. Meanwhile choice accuracy across 32 possible options reached 36%.
The model is also designed to flag uncertain decisions as requiring human review rather than executing a tool blindly. TokenAI calls this built-in abstention behaviour ‘confidence-aware routing’.
Running on CUDA hardware, Neo processes roughly 10,241 decisions per second with individual decision latency under 0.1 milliseconds, reflecting its compact size relative to general-purpose language models.
The model, training code, dataset and documentation are published openly on GitHub and Hugging Face under the TokenAI Neo Model License, which prohibits commercial, client-facing or production use without written permission.
Neo is the sixth significant code release from TokenAI in 2026.
ZOOM OUT - Neo caps a string of 2026 model releases from the Alexandria-based TokenAI. The year started with Horus 1.0-4B in April, a fully open-source model that scored 88 percent on the MMLU benchmark, ahead of larger rivals like Qwen 3.5-4B and Llama 3.1-8B. July brought Horus Hiero, built to read Ancient Egyptian hieroglyphics alongside modern Arabic dialects, followed by Horus Cyber Nano for cybersecurity work. September alone saw three releases: the Ein-Horus Arabic pretraining corpus, Horus Taleeq 0.2B Base, and now Neo. Six releases in nine months, all open or open-weight, is an unusually fast cycle for a brand new AI startup.
[Written and edited with the assistance of AI]
Source: Token AI
LINKS
Neo model page (Token AI)
Neo code (Hugging Face)
Neo dataset (Hugging Face)
Neo complete repository (Github)
Read more about 2026 TokenAI model releases:
TokenAI open-sources Horus Taleeq Arabic language model (Middle East AI News)
TokenAI releases Horus Cyber Nano 1.0 (Middle East AI News)
TokenAI to release Horus cybersecurity model (Middle East AI News)
TokenAI’s Horus Hiero multimodal AI reads hieroglyphics (Middle East AI News)
Egyptian open-source LLM Horus punches above its weight (Middle East AI News)