Skip to content
Atmane Ayoub
← Notes

Note · 28 Sept 2026 · 1 min read

I released Jev-AR, an Arabic-first routing model

322M parameters, one forward pass, routes supplied at request time — for agents and RAG in MSA, Emirati, Saudi and mixed Arabic-English.

I built Jev-AR — a 322M-parameter Arabic-first decision model for routing in AI agents and RAG.

Instead of training a fixed intent classifier, you give Jev-AR the possible routes at request time: agent tools, RAG knowledge bases, workflows, support queues or application actions. It scores them in one forward pass and returns a probability distribution over the choices. New route list? No retraining.

It's built for MSA, Emirati Arabic, Saudi Arabic and Arabic-English mixed text.

The headline result: 83.4% on 4,974 held-out Gulf-Arabic queries over 60 routes — 6.6 points behind a 27B LLM — at about 10 ms p50 on a laptop GPU. No LLM generation, no remote API, no per-call inference cost.

It isn't perfect: the training data is synthetic, native-speaker validation is still needed, and completely unseen route semantics remain the area I want to improve most.

The model and benchmark are public for anyone who wants to test it, break it or build on it.

→ The full write-up: architecture, training and results