A startup inside Midcore
Train on the laptop you have. Cloud-burst when you must.
A PyTorch sidecar with eight training paths and automatic GPU probing across NVIDIA, Apple Silicon, and AMD. The data never leaves the device unless you tell it to. The bill stops when the laptop closes.
Eight training paths
Behaviour cloning to QLoRA, one runner.
Behaviour Cloning · PPO · SAC · LoRA · QLoRA · Distillation · Full ≤3B fine-tune · Custom recipes
GPU probe
Detects the silicon you actually have.
NVIDIA, Apple Silicon, AMD. The probe runs once on first launch, picks the right backend, and tells you up front what your machine can hold. Honest about its limits — the LoRA τ₀ world-model fits on RTX 4080+; the full τ₀ stays cloud because nothing else has 70 GB of VRAM.
Local-only mode
A switch that means it.
Flip the local-only switch and every outbound midcore.ai call — telemetry, license refresh, compute catalog, provider proxy — is gated. Direct-vendor LLM calls and local compute runners still work. Your data does not leave the device.