Developed by Moroccan software engineer Rachid Nichan, Awal-LM is designed specifically for Tamazight and can generate and respond to prompts written in Tifinagh. The project is available through GitHub and Hugging Face, with a live version also accessible online.
The name “Awal” means “word” or “speech” in Tamazight. The project describes itself as an open-source effort focused on improving digital tools for the Amazigh language.
Awal-LM has gone through several stages of development. An earlier version of the project was presented as a lightweight neural-network engine built from scratch in C++, without relying on machine-learning frameworks such as PyTorch or TensorFlow. It included a custom tokenizer designed to handle the 33 Tifinagh characters standardized by Morocco’s Royal Institute of Amazigh Culture.
Read also : AI decodes 217-year-old Napoleon letter in six hours
The current v1.8 model takes a different technical approach. It is based on Google’s Gemma-2-2B architecture and was fine-tuned with Unsloth using a curated Tamazight dataset. The model uses LoRA adapters and is currently described by its developer as experimental and still under development.
The model’s Hugging Face page lists Central Atlas Tamazight among its supported language information and identifies Google’s Gemma-2-2B as its base model.
Designed around Tifinagh
One of the project’s central features is its focus on native Tifinagh input rather than treating Tamazight simply as another language added to a general-purpose chatbot.
Awal-LM includes Tifinagh-specific tokenization and supports interactions directly in the script. Its online interface also provides a virtual Tifinagh keyboard, allowing users to test the model without having to install a separate keyboard layout.
The developer says the model was trained on curated Tamazight material, including IRCAM-related data, literature and custom instructions. The latest version is intended to follow simple instructions and answer questions in Tamazight.
An open-source project for a low-resource language
Tamazight remains a relatively underrepresented language in mainstream AI systems compared with languages such as English, French and Arabic. Projects like Awal-LM seek to address that gap by developing language-specific datasets, tokenizers and models.
Read also : Meta launches personal AI agent that can send emails, make payments and book travel
Awal-LM is also listed in the Tamazight-NLP collection on Hugging Face alongside other language, speech and text resources developed for Amazigh varieties.
Nichan has also worked on other Amazigh-focused technology, including an application designed to help users learn Amazigh through Tifinagh and interactive quizzes.
The project remains experimental, but its open-source approach allows developers and researchers to inspect the model, build applications around it and contribute additional Tamazight data. The Awal-LM project specifically calls for contributions to dataset expansion, human feedback and applications using its API.
