Still Light
Outside

For the 90s kid who grew up in the 2000s & 2010s: screen names and away messages, burned CDs with marker handwriting, bikes dropped on front lawns, being called home when the streetlights came on, wired headphones drowning out road noise.

They're around thirty now. Summer is ending — not their last summer, but the kind of August evening where you suddenly notice you've been happy, and noticing already feels like missing it.

The title is our thank-you to that feeling. Still light outside — telling it that the sunset is beautiful, that we look up for the stars that come out at night, and that we know: soon there will be a new dawn.

MiniMax released Music 3 as an open-weight model, and we pulled the repository the day it landed. Day one assumed CUDA: no MLX, no MPS, no road laid across Apple silicon. So we laid one ourselves, first teaching the released pipeline to survive on a 64 GB M1 Max, then carrying that work forward until an Apple-native implementation was merged upstream. The first long local performance asked for 342 seconds of music and returned 237.47 seconds after 7,414.95 seconds of inference — two hours, three minutes, and thirty-five seconds — with a 71.1 GB process footprint and roughly 41 GB of cumulative swap-out writes.

Then we shaped the pipeline for the machine instead of asking the machine to pretend it was a server rack. Staged loading cut that footprint from 71.1 GB to 31.13 GB and stopped render-caused swap-outs; the finished hybrid path held about 22.03 GB. On the same 15-second, 30-step gate, wall time fell from 346.8 seconds in staged PyTorch/MPS to 244.8 in native MLX, 237.9 after restricting the semantic head from 200,000 rows to 16,385, 199.5 with component-specific precision, and finally 178.4 — 2:58.4 — with batched guidance.

Every one of those numbers had to answer to a pair of ears. A 20-step shortcut cut 217.5 seconds to 178.3 but left audible distortion; an RVQ depth cache cut 237.9 to 217.5 and came back off-key and garbled. Both were struck from the pressing. Batched guidance stayed because it saved 10.6 percent of wall time, measured 94.47 dB difference SNR against the reference, and still passed a human listening gate. A metric cannot hear. A change survived only when the render got faster and a person still loved what came out; the failures stayed in the ledger.

The long-form program climbed by guarded rungs: 15, 60, 120, then a 180-second ceiling, with checkpoints after the semantic pass and every acoustic window. The approved E-flat take passed its 60-second gate in 13:57.4, continued to 120 seconds in 28:22.1, and finally found a natural ending at 172.70 seconds in 41:32.4. It held 101.12 BPM, wrote zero swap-outs, and could resume after interruption without changing a byte of the finished audio.

These thirteen tracks were rendered through the hosted service while that local press was being built. And now the machine on the desk can do it too: music made completely offline, in a bedroom, from a feeling and a few words. That is the Baiograph promise made tangible — a tool from the future that can carry those feelings into any genre, with the craft of singers, bands, and sound engineers close at hand.