Fig 1Charizard · local AI rigTurn the knob · 1/2/3 or ←/→dsv4.1-flash · 2 cards · 105 tok/s · 1M ctx · 42 beads
The brass key: the gateway's lock, on 127.0.0.1:12434Two RTX PRO 6000 Blackwell cards, 96 GB each, held to 450 WToken beads on the rail: one bead is 25,000 tokens of contextThe Tailscale cable to the laptop, 100.The rig on the bench. The laptop is the terminal: it reaches the machine over Tailscale at 100., and the machine answers only through the keyed gateway on 127.0.0.1:12434, whose brass key lives in ~/.local/state/omarchy/local-ai/gateway.key. Turn the selector knob to load a recipe: the cartridge swaps in the Docker engine housing, the cards it uses light, and the token beads set off from the gateway along the rail to the cards and come back down the cable to the laptop. Bead count is the model's context, one bead to 25,000 tokens; bead speed is its decode rate — Nemotron 3.5 Lightning 385 tok/s on one card at 200K, Qwen3.8-27B 184 tok/s on one card at 200K, DeepSeek-V4.1-Flash 105 tok/s tensor-parallel across both cards at 1M. A one-card model takes the top card, the first free one, which is also the card the display is on; a two-card model fills the rack. The UPS carries the machine and its cord ends at the bench outlet.