The Nomeda Ecosystem
Arabic AI, From the Ground Up
From the Fattah-Orch family and Arabic tokenizers to open datasets — every layer of our stack is designed for Arabic-first AI.
Fattah-Orch Family
A suite of lightweight orchestrator models (XS 0.6B to L 8B) that decompose Arabic and English coding requests into structured JSON task graphs. Built on Qwen3, fine-tuned on the first Egyptian Arabic software task dataset — each model routes directly to downstream coding agents.
Arabic-First Tokenizers
Specialized tokenizers for Modern Standard Arabic and Egyptian Arabic — Nomeda-MSA-64K and Nomeda-Egyptian-16K. Designed to capture the full linguistic spectrum from formal text to everyday dialect with high fidelity and low compression loss.
Open Arabic Datasets
We release training data openly to accelerate Arabic AI research. The Fattah Orchestrator Dataset is the first Egyptian Arabic software task decomposition dataset, and Hindawi Arabic Sections provides 52K+ labeled book passages for general Arabic NLP.
Enterprise-Grade Deployment
Designed with production in mind. Our models can run on any target device — from CPU to GPU — and deploy on-premise or in secure cloud environments, giving organizations full control over their AI infrastructure.
Our Approach
What Drives Us
Arabic-Native by Design
Built from the ground up for Arabic and its dialects. Our models understand Egyptian Arabic and MSA natively — not as a translation afterthought — making them more accurate and contextually aware.
Open by Default
Model weights, tokenizer configurations, and datasets are publicly available on HuggingFace. We believe Arabic AI research progresses fastest when the community can build on shared foundations.
Practical by Design
From lightweight 0.6B models running on any CPU to 8B models for GPU inference — our ecosystem is built for real-world deployment, not just benchmarks. Task routing and model specialization keep costs low.
Explore Our Models
Check out our open-source models, tokenizers, and datasets on HuggingFace.