Aragt
Contact
Open Source

Knowledge accelerates when it's shared.

We contribute to the global AI ecosystem by releasing models, code, frameworks, benchmarks, and educational material,openly, whenever we can.

Why open

Open science has driven most of the important progress in artificial intelligence. Every generation of researchers stands on open code, open datasets, and shared benchmarks.

We open-source what we can, especially when it lowers the barrier for developers, researchers, and institutions in places where AI has historically been an import rather than an export.

We do this responsibly,with model cards, evaluations, and usage-policy guidance. Open does not mean unaccountable.

What we release

Six kinds of open contributions.

  • Models

    Open-weight language, multimodal, and speech models trained by our research teams.

    Weights + configs + eval reports

  • Research Code

    Reference implementations for the ideas behind our papers.

    Runnable + reproducible

  • Frameworks

    Libraries for training, serving, and evaluating modern AI systems.

    Batteries included

  • SDKs

    Typed SDKs and CLIs that make the Aragt platform easy to use.

    TS · Python · Go

  • Benchmarks

    New benchmarks,especially for capabilities that don't have good ones yet.

    Public leaderboards

  • Datasets

    Curated datasets, released where legally and ethically appropriate.

    Data cards + licenses
Releases

What we're working on in the open.

A running list of releases in progress. Follow along and contribute.

Get release updates →
  1. Model
    aragt-core-base

    Open-weight base language model, permissively licensed.

    Coming soon
  2. Model
    aragt-african-speech

    Multilingual ASR + TTS for African languages.

    In progress
  3. Framework
    aragt-agents

    Framework for building, evaluating, and running agents.

    In progress
  4. Framework
    aragt-evals

    Batteries-included evaluation harness for LLMs and agents.

    Coming soon
  5. Benchmark
    african-language-bench

    Community benchmark for LLMs across major African languages.

    Coming soon
Policy

Our default open-source license is permissive. Model weights are released with a use policy that prohibits harmful applications.

Datasets are released only where we have a clear legal and ethical basis to do so, with full data cards and attribution.