Roadmap

Our software stack includes a Mac launcher app, a cross platform SDK based on Modelfile¹, and a decentralized registry built with Swift. The Modelfile serves as a declarative blueprint for creating and sharing open models, enabling the generation of Modelpacks for multiple platforms from a single source. Our goal is to develop this in collaboration with the community and establish it as the standard for customizing ML models.

Our technical approach is guided by the insights of The Bitter Lesson² and inspired by the vision in Inventing the Future³. We are building a high-end general system based on the hypothesis that abundant compute will continue to become more accessible on local machines. We're starting with Apple platforms using MLX, as their unified memory architecture and strong platform security make them well-suited for running private local models. We chose Swift to ensure compatibility across platforms, including iOS devices.

The project consists of four design choices⁴ that support and reinforce one another:

1. Device-anchored identity with keyless ops: Clients must be provisioned through the device key chain and cannot access the registry by identity alone⁵. Only after an identity is verified and linked to the device key can it enable keyless operations⁶ for seamless private training on CI/CD pipelines.
2. Immutable model builds: Every build is version-locked and reproducible, ensuring consistency and reliability across updates and platforms.
3. Content-hashed model layers: Models are stored and referenced by cryptographic hashes of their layers, guaranteeing integrity and enabling efficient deduplication and sharing.
4. Verifiable transparency and attestations: Every signing and build event is recorded in an append-only transparency log, producing cryptographic attestations that can be independently verified. This ensures accountability, prevents hidden modifications, and provides an auditable history of model provenance across devices and registries.

As part of our research on private software personalization, we are exploring research in Trusted Execution Environments (TEEs), Per-Layer Embeddings (PLE) with offloading to flash storage, runtime LoRA generation with hypernetworks, and reinforcement learning techniques such as GRPO.

We are seeking design partners for training workloads that align with our goal of ensuring a verifiable privacy perimeter. If you're interested in partnering, please reach out to us at feynon@tiles.run.