ej

ej launches with a very specific job: take some state, take a set of decision questions, and return a probability distribution for each one in a single pass. That’s a narrower aim than general classification, and it’s the whole point. The model is built for questions whose answer choices can be defined at runtime, which makes it more flexible than a fixed-label classifier without drifting into full-blown text generation.

The project says it handles choice, yes/no, and ordered scores. It runs on CPU, so it does not depend on a heavier setup to be useful, and it does not generate tokens at all. That keeps the model in a different category from the usual chatbot machinery. Smaller, simpler, and easier to slot into a decision pipeline.

The distribution is packaged as one 11.4 MB model file. That is tiny enough to get attention on its own, but the repo is doing more than shipping weights. It also includes training, evaluation, adaptation, packed weights, and a benchmark harness, which suggests the team wants people to inspect, reproduce, and bend the system rather than just download it and hope for the best.

This is open-source AI pitched around a practical constraint, not a grand claim. The launch tags it as Artificial Intelligence and GitHub, and the framing fits: a focused model for structured decisions, not a universal brain in a box.