About
Compiwer AI is an independent AI lab and part of Compiwer Group. We build open models, an evaluation benchmark, and software for running models locally.
Most capable AI systems are built by a few large organizations and work best in a few languages. We think small labs can do useful work by adapting open models to needs that are underserved, and by publishing everything they make. The Mtrini model cards list Salé, Morocco as their origin.
How we work
- Open weights
- Every model is published on Hugging Face with its weights, usually as an adapter, merged weights and a GGUF build.
- Built on public foundations
- Mtrini models are fine-tunes of public base models from Qwen, Google and HiDream. Each model card names its base model, whose license and limitations still apply.
- Documented training
- Model cards list training data, settings and training-run metrics. Training loss is reported as what it is, not as a benchmark score.
- Local inference
- Quantized GGUF releases, the Compiwer SDK and Mtrini Workspace are aimed at running models on ordinary hardware, offline.
- Underserved languages
- Moroccan Darija is spoken by millions of people and rarely evaluated. It is part of our training data and has its own dimension in RMCMMK-Bench.
Where things stand
Our releases are early. Several are marked as experiments, research previews or development checkpoints on their model cards. Benchmark results are published only through our benchmarks: ATU-1.1 has official standings that include our own models alongside others, and RMCMMK-Bench has no verified results in its repository yet.