Our work on this track leads to training and fine-tuning small open models that run locally. We fine-tune, distill, and quantize them so they fit a laptop, a desktop GPU, or Apple Silicon. You keep the weights. It can run fully offline on your hardware.
No price is listed for this work. Firms that want the deployed assistant: the $3,000 Private AI Enterprise Architecture Sprint is a separate track.
The result is a model you run yourself. We don't host it.
We train an open model on examples of a defined task: classification, extraction, drafting, routing, or another job you describe.
When the capable model is too large for the machine, we train a smaller open model to follow it on that task.
We reduce the numeric precision of the weights so the model uses less memory and can run on consumer hardware.
You receive the weights and short notes on how to run them. We don't keep a hosted copy for you to call.
We say which of these steps a job actually needs before work starts. Some jobs are only quantization. Some are a fine-tune plus quantization.
This track is a model sized for consumer hardware. It is not a firm-wide assistant deployment.
You want inference in a product or a script you control, without sending each call to a hosted API.
Captions, cleanup, a house style, or another job you do often, running on the computer you already use.
A few people, one or two machines, and a task specific enough that a small open model can do it.
Prompts and files stay on your hardware when you run the model there. It can run fully offline on your hardware.
We agree the machine and the method before training. Timing is part of that conversation. Contact us.
You tell us what the model should do and where it has to run: a laptop, a desktop GPU, or Apple Silicon. If you are not sure it will fit, we say so.
We choose an open model and whether to fine-tune, distill, quantize, or combine those, sized for that hardware.
When you have examples, we train on part of them and check the rest, held out from training. When you don't, we agree on a public dataset, or we stop before training.
You get the weights. Where they run is up to you. It can run fully offline on your hardware.
If a job uses your examples, we use them to train and to check the model. We don't publish your data. Jobs we have already published use public data. They are not customer results, and they are not a speed claim for your machine. See fine-tuning.
Consumer hardware means machines people buy for themselves. Fit depends on the model and the machine. We have not published speed numbers for this.
Only when the model is small enough for that laptop. We don't treat every laptop as the same machine.
We size the model for the card you name. We don't assume a data-center GPU.
Same rule: the weights have to fit the machine you will actually run.
No price is listed for this work. Contact us and we will reply about scope. Or book a call.
650-318-8808 · jordan@booleinference.com
For a firm that needs a private assistant on infrastructure it controls, the $3,000 Private AI Enterprise Architecture Sprint is unchanged: live within 48 hours of kickoff and data access, or a full refund. Apply for the sprint.