Roadmap
A generational project, laid out in phases. This is the order things arrive in, not a calendar of dates.
NexaCore OS is a multi-year build, and we plan it that way. The phases below are indicative and high-level: they show the sequence in which capabilities land, not when. Detailed milestones are set at the start of each phase. For the current line between what works and what is still a specification, see where the project stands and the devlog.
The sequence
Each layer is built, hardened, and — where it crosses a trust boundary — externally audited before the next one depends on it.
The Rust microkernel boots bare-metal on x86-64 and runs on real hardware: memory, scheduling, typed message-passing, and the capability primitives, with user-space drivers for NVMe, VirtIO-net, and e1000e, and a from-scratch AI inference pipeline running from a Ring 3 process. This is the trusted base everything else stands on. It gets an external security audit before it carries real workloads.
The AI Runtime Service, exposed through system calls, with a tensor layer that dispatches to CPU, GPU, and NPU. Inference runs entirely on the local device by default. The first encrypted-by-default data types land in the OS API, alongside the local tokenization service that strips identifiers before anything is shared. A signed reference model, verified before it loads.
Your own devices, working together. NexaCore OS finds other instances on your network, attests them, and connects over mutual TLS, then splits larger models across them. Nothing leaves machines you own.
The opt-in peer-to-peer mesh: distributed discovery, a QUIC and Noise transport with mandatory mutual attestation, TEE-only encryption envelopes, a compliance proof on every message, a signed compute-credit ledger, reputation, and onion routing for sensitive work. Mixture-of-experts models distribute across peers, and results are checked by redundant execution. This is the v1.0 release: inference across all tiers, on x86-64 with hardware trusted execution. It ships after a comprehensive external audit.
Apple Silicon with Secure Enclave attestation, and performance work driven by real v1 usage on real machines.
Training across the mesh, not just inference: secure aggregation of gradients, differential-privacy noise, relaxed-sync training between peers, and a community model registry where experts can be trained, forked, and merged.
Honest caveats
This roadmap is indicative, not a promise of dates. NexaCore OS is planned in horizons of years, and some steps depend on the security foundations holding and on the resources to do them properly. Where a layer isn't ready, we would rather move it than ship it early.
At each phase we evaluate openly whether to continue, adapt, or stop, and we say so in the devlog. A roadmap you can hold us to is worth more than one that only ever points up and to the right.
Go deeper
The technical deep dive covers the microkernel and its system-call ABI, the AI runtime, the cryptographic privacy model, and ncScript. The devlog tracks progress as it happens.