Inference setup.

Every product runs its agents on APG, and APG draws inference from one of three tiers. Pick the tier that matches your privacy and cost posture, then set it with a config edit.

Read the infrastructure overview →

Three ways to run it

Where the model runs.

Your agents run on your own machine. The model behind them can run through us, on your own key, or on a computer you own.

Through us: Metered and itemized. Not open yet.
Your own key: You hold the key. No inference margin from us.
On your machine: Nothing leaves your network.

Where the model runs

Your agentsSolutions FactoryModel provider

Where it goes: Your agents send each request to us. We pass it to the model provider.
Who sends the bill: You pay us for what you use, itemized like a utility bill.
Your data: Your records stay in your own database. Only the request itself passes through.
Status: Not open yet. What is kept, and for how long, is stated in writing before it opens.

Tier 1 · SF default

Route through Solutions Factory.

Tier 1 is coming soon and cannot be switched on yet. When it opens, APG will proxy inference through Solutions Factory and we will bill you per token, itemized, with nothing for you to procure. Until then, set up Tier 2 or Tier 3 below. The commands that will enable it are shown for reference.

$ apg config set inference.tier sf-default
$ apg config set inference.sf_token $SF_INFRA_TOKEN
inference: sf-default · metered + itemized · proxy designed for zero data retention, to be confirmed before launch

Tier 2 · BYOK

Bring your own model key.

Point APG at your own model key. Your model provider bills you directly, and Solutions Factory takes no inference margin.

$ apg config set inference.tier byok
$ apg config set inference.api_key $MODEL_API_KEY
inference: byok · billing direct to your model provider · SF margin: none

Tier 3 · Sovereign

Self-host the model on your GPU.

Install a local model runner, pull a 7B-or-larger model at Q4 quantization or better, and point APG at it. No prompt content leaves your network. A single GPU with 12 to 16 GB of VRAM handles an 8B model at Q4.

$ # download a 7B-or-larger local model, Q4 quantization or better
$ apg config set inference.tier sovereign
$ apg config set inference.local_model <your-local-model>
inference: sovereign · local model · nothing leaves your network

We are still writing deeper guidance on model selection, quantization trade-offs and multi-GPU setups.