Classic 2D · All connections	Classic 2D · Selected neuron connections
Coaxial Rings 3D · All connections	Coaxial Rings 3D · Selected neuron connections
Spiral Tokens 3D · All connections	Spiral Tokens 3D · Selected neuron connections

Explore the whole network, then focus on a single neuron. Click any image to open the full-resolution image.

MiaIA Studio

An interactive development environment for understanding neural networks.

VISUALIZE  |  EXPERIMENT  |  INSPECT  |  DEBUG

Website  ·  Documentation  ·  Roadmap  ·  MPL 2.0

Understand the network, not just the output

MiaIA makes neural networks observable. Build or import a model, execute it, inspect activations and gradients, follow parameter updates, and advance training one mathematical phase at a time.

The MiaIA Studio 0.1 Alpha release includes:

• a C++20 neural-network Engine and public SDK;
• direct inference, evaluation, observable backpropagation, and atomic SGD training;
• controlled foreground and background sessions with navigable history;
• phase-by-phase training debug with candidate-state inspection and rollback;
• immutable forward and backward gradient-flow traces with graphical playback;
• multi-model projects built from isolated model contexts, including clean experiment forks with copied networks, datasets, training configurations, and breakpoints;
• immutable comparison of two model contexts with topology compatibility, ranked parameter deltas, and 2D/3D visual overlays;
• model checkpoints with inspection, comparison, transactional restore, and .mai persistence;
• bounded neuron and connection relationship inspection with exact topology counts;
• ONNX model interchange and numeric CSV datasets;
• one shared command processor for the terminal and Unreal clients;
• interactive 2D and 3D topology views in MiaIA Studio;
• a renderer-neutral StudioCore application layer and Windows standalone host.

Project status — 0.1 Alpha: the mathematical and application foundations, versioned .mai v2 multi-context persistence with v1 migration, execution traces, diagnostics, and model checkpoints are implemented and tested. Persisting session history and visualization layouts remains planned work. APIs and supported file-format behavior may evolve during the alpha series.

Alpha limitations

• The native engine currently focuses on observable feed-forward networks and the activations documented by the project.
• ONNX import and export support the documented dense subset, not arbitrary ONNX graphs or operators.
• Built-in optimization currently focuses on mean squared error and stochastic gradient descent.
• .mai version 3 persists deterministic sample-order configuration but not transient training progress, session history, annotations, or visualization layout; versions 1 and 2 remain readable.
• Live model, dataset, training, and checkpoint state is process-local; separate executables do not share one running session.
• The packaged application is currently verified for Windows x64. Other platforms and solution configurations are not release targets yet.
• Alpha APIs and workflows can change; preserve important interoperable models through ONNX exports where supported.

Download the Windows app directly from the Microsoft Store. Visit www.nonop.biz/miaia for official project information. The corresponding source code is maintained in this repository.

Documentation

• Documentation overview
• Architecture
• History
• Roadmap
• MiaIA Studio

License

Original MiaIA source code is available under the Mozilla Public License 2.0. Third-party components, Unreal Engine, and MiaIA branding remain subject to their respective terms. See Licensing, third-party notices, and trademark policy.

Copyright 2026 Agostino Mosillo.

The repository is currently author-led and does not accept unsolicited code contributions. Bug reports and focused feedback are welcome as described in CONTRIBUTING.md.