I built a record label
that runs itself.

This isn't one model wearing hats. It's a deliberate hybrid — frontier models where judgment is worth paying for, an open-weight model at home for everything that has to run continuously and never leave.

The split is the point. Frontier models are extraordinary and they are somebody else's computer. So they get the work that benefits from them — design, judgment, code — under my hand, session by session. Everything that runs unattended, and everything that touches the catalog, runs on Ficus, at home.

Ficus thinks with gpt-oss-120b — OpenAI's own open-weight release, Apache 2.0. Not a derivative, not a distillation. The published weights, on my disk.

• 117 billion total parameters, 5.1 billion active per token. It is a mixture-of-experts model, and that ratio is the entire reason this is possible on a desk.
• What that means. A dense model runs every parameter for every token — 117B of knowledge would cost 117B of arithmetic, every word. An MoE splits the network into many specialised experts and puts a router in front. For each token the router wakes only a handful.
• So it carries the knowledge of a 117B model and pays the compute bill of a 5.1B one. Roughly twenty-three times more knowledge than compute. That's not a compromise — it's how the frontier got small enough to come home.
• MXFP4 quantisation on the expert weights. Four-bit precision where the experts live, which is what lets 117B parameters fit in memory at all.
• It normally wants an 80 GB datacenter GPU — an H100 or an MI300X. Mine runs on a 240-watt appliance next to a monitor.
• A vision model sits beside it — qwen3-vl — so the system can look at video, images and screens rather than only read text.
• And a memory that rebuilds nightly. A retrieval store re-indexed at 04:30 every morning, so the agents wake up knowing what happened yesterday.

A model with no memory is a very fast stranger. So the agents share one — a local retrieval store, 4.7 GB of it, living on the same box as the model that reads it.

• It's tiered by what kind of knowledge it is. Four layers, and retrieval knows which is which: M1 — what each agent is, its contracts and levels. M2 — what the system learned, written the moment it's learned. M3 — standing doctrine, the rules everything runs under. M4 — general long-term memory. A question about how something works and a question about what happened last Tuesday do not go to the same shelf.
• Two embedding models, not one. nomic-embed and bge both run over the corpus, because different text answers to different embeddings and a single one leaves recall on the table.
• Rebuilt nightly at 04:30, from the live system. Not a snapshot somebody took once. Every morning it re-derives from what's actually on disk, so the agents wake up knowing the current state rather than a remembered one.
• The previous generation is kept, not overwritten. Last night's corpus sits beside this morning's. A bad refresh can't quietly destroy the record — it can only sit next to it.
• There's a quarantine. Documents that shouldn't be indexed get isolated instead of embedded. The store having an opinion about what doesn't belong in it is the difference between a memory and a landfill.
• It's served, not imported. A retrieval service runs on the box with its own endpoint, so every agent asks the same memory the same way and nobody keeps a private copy that drifts.
• And it never leaves. The entire corpus — contracts, canon, what the label knows about itself — is embedded locally and queried locally. No document in it has ever been uploaded to be indexed by anyone.

An NVIDIA DGX Spark. About a quadrillion calculations a second on a desk.

• GB10 Grace Blackwell superchip — up to 1 petaFLOP of AI compute at FP4.
• 128 GB of coherent unified memory at 273 GB/s. CPU and GPU share one pool — nothing is copied back and forth across a bus. That coherence is why a 120-billion-parameter model fits here and not on a gaming card with more raw speed.
• 20 Arm cores — ten Cortex-X925 and ten Cortex-A725.
• 4 TB of self-encrypting NVMe.
• 240 watts at the wall. Less than a hair dryer.
• Rated for models up to 200 billion parameters locally. There is headroom.

Same headline number, different arithmetic — Roadrunner's petaflop was double-precision, mine is the low-precision math AI inference runs on. I'd rather say that than have you find it. Eighteen years ago a petaflop of anything took a building, two megawatts and a hundred million dollars.

The three minds don't do the work themselves. They run a company, and it's split the way a label is actually split — a corporate side that keeps the business alive, and an artistic side that makes the thing the business exists for. Every entity on both sides carries a written contract naming exactly which tools it may touch.

Each officer owns departments; departments own workers. Real names, real jobs, real reporting lines.

• Margot keeps the books. Silas watches the bank. Both sit in Accounting, under Helena.
• Corey writes and answers mail, in Communications, under Cassius. He wrote the last three emails that left this company.
• Apollo runs paid promotion. Enzo enriches data. Two GRMEC agents handle city and industry targeting.
• Vera is the key master — credential custody, the one that holds what the others may not see. Argus watches the runtime.
• And six vitals in the Command Center — heartbeat, scheduler, logging, usage, backup, and a watcher that unit-tests the shared toolbox every fifteen minutes. The system's job is to notice when the system is broken.

Not personas on one account. Each has its own identity, its own voice, its own catalog, and its own Instagram, YouTube, TikTok and Facebook. They release music. They answer their own mail.

• Each one is drilled against its own rubric. There's a testing pipeline that scores whether a chump still sounds like itself — voice adherence, measured, not assumed.
• They are grounded before they speak. A drafter assembles what's actually on disk — real assets, real numbers — and the chump has to write from that or refuse. An artist that invents a song it never made is worse than one that says nothing.
• Records watches and cannot speak. One tool, read-only. The label observes; the artists talk. That asymmetry is deliberate.
• J. Carmichael is a letter-writer, not a poster. A different tool family entirely — correspondence on the real domain rather than social posting. Not every artist needs the same hands.

The system doesn't stop at software. There's a Bambu Lab P1S on the network, and the operators drive it end to end.

• It runs on the LAN, not the vendor cloud. Credentials are vaulted alongside everything else, and jobs go over implicit FTPS on port 990 straight to the machine. Nothing about what I make is uploaded to anyone.
• The operators handle the whole chain. Find the model, check it against the loaded filament, slice it, verify the estimate, send it. This morning I asked for a bedside phone holder for my mom. The operator picked the model, confirmed the assigned spool wasn't the empty white one, sliced it — 1h23m, 54.6 g, one dollar forty-two — and sent it to the plate.
• Status probes watch it — the same way the rest of the fleet is watched. If it stalls, that shows up where everything else shows up.
• Honestly stated: the printer is a rail today, not yet an agent with its own contract and reporting line. The wiring, the vault entry and the probes are real; the entity around them is the next build. I'd rather tell you where the edge is than let you find it.

Anything that touches money or reaches the public is gated behind a token that is the cryptographic fingerprint of the exact bytes I approved. Change one character after I sign off and the gate refuses. Verifiable rather than trusted — the only version of this worth building.