Cinema + labs + WebGL. The labeled map, Hub canvas, field sketch, and every live engine. Use when this is the app that owns the GPU on the machine.Every feature, no graphics. Same engines, same API, same honesty grammar. No WebGL, no custom cursor, no Three.js. For when other apps already hold the GPU, or after learning loops when you only need the work.Terminal + one API out. Every live verb as a command and as HTTP. TUCE agents load skill.md and call /api/v1. No cinema overhead. Occupancy stays nominated | live-bound | unavailable.
Lab · 15 · Illustrative serving math
KV & quant plate
Context × params × batch for a KV sketch, plus Q2–FP16 weight bytes. Labeled illustrative. Not a serving SLA.
NCM 1.19.0 · Hub 2.2.0 · invented_metrics = false
illustrative
Illustrative serving math
Not a serving SLA. Tokens/sec are not published.
KV sketch · FP16 · GQA 8×128 × 32 layers
1.000 GiB
2 × layers × kvHeads × headDim × ctx × batch × 2 bytes. Params (B) scales the weight table below, not the KV tensor — KV does not contain the MLP weights.
← Spoke budgetAdd illustrative GiB for a hot set. Download-only is not running. Tokens/sec are not published.All labsEvery this-tab tool. Real DPLL, occupancy, adapters, SAT. Not the Hub runtime.WASM SAT race →Same CNF, three engines, this-tab timings. SIMD is a detect lamp, not a claimed kernel. Not a benchmark publication.