Coming Soon
Every byte for inference.
Kelvin is a minimal, immutable Linux distribution that turns Apple Silicon hardware into a private AI inference appliance. Plug in, run models. Nothing else.
macOS holds 2-5 GB of non-evictable wired memory hostage and caps GPU memory at 75%. Kelvin reclaims all of it. The result: fit models and context lengths that simply don't fit under macOS, on hardware you already own. Need more? Buy a second Mac — Kelvin clusters them automatically over Thunderbolt, zero config, and runs models too big for any single machine.
The product
- Run bigger models — no macOS RAM tax, no GPU-memory ceiling. Every GB goes to the model.
- Clustering just happens — tensor-parallel across nodes, automatic peer discovery, zero setup.
- Hot/cold tiering — GPU for the hot layers, CPU/AMX for the cold. Spill-over degrades gracefully, not catastrophically.
- Install and use — no OS to manage, no packages to install, no shell to break. The complexity is our job to hide.
By design
- Immutable, read-only root (dm-verity). Tamper a single block and boot fails.
- Verified boot chain sealed to the Secure Enclave.
- No user shell. No package manager. All admin via one audited control plane.
- Signed, licensed workloads only. The admission rule has no exceptions.
- Always-on, always-connected. Zero-touch provisioning + remote observability.
Built on Linux, not from scratch. For inference, the OS isn't on the performance-critical path — the silicon is. We keep the mature GPU compiler, networking stack, and memory subsystem. We hack the memory layout where it matters (huge pages, contiguous reservations, TLB-friendly streaming) and spend our effort on what actually moves the needle: the engine and clustering.