Pick your tier.
Run it locally.

The old constraint: fit the entire model in VRAM. A 16 GB consumer GPU can't load a Q4-quantized 13B model. It just doesn't fit. For a long time that felt like the ceiling. Your options were a $10K+ workstation GPU or a cloud API, and most people quietly chose the API.

Apple Silicon changed this in a way I don't think has been fully absorbed. GPU and CPU share one memory pool. A 16 GB Mac mini loads a Q4 8B model (~5 GB) with room left for the OS and other apps. There's no VRAM ceiling.

Every model listed here is open-weight, publicly released weights, no API keys, no telemetry, no vendor agreement. The weights you download today run the same way in five years. Air-gapped by default.

I get this question a lot: does local pay for itself in API savings? Honestly, not on token costs alone. At $0.01 per interaction, you're looking at years before hardware pencils out. That's not the argument and I won't pretend it is.

The real cost is what I'd call compliance overhead: the vendor due diligence, the BAA negotiation, the conversation with your malpractice carrier about someone else's breach notification process. A local machine removes the third party entirely, not cheaper per token, but simpler in a way that actually matters to a solo practice.

There's also a cost nobody thinks about until they see the invoice: runaway API calls. Cloud models are metered at both ends: GPT-5.5 runs $5 per million input tokens, $30 per million output. GPT-5.5 Pro is $30 in, $180 out. Most of the time that's manageable. But agentic workflows that loop, integration bugs, or a script left running overnight don't stop billing because something went wrong. The meter runs regardless. A $20/month habit becomes a $1,200 charge before anyone notices. With local inference, there is no meter. The model runs or it doesn't. Nothing accumulates.

And then there's the subscription stack. Most people I talk to aren't paying for one AI service. They're paying for two or three. ChatGPT Plus at $20/month, Claude Pro at $20/month, maybe Copilot on top of that. If you upgrade to ChatGPT Pro it's $200/month. Claude Max starts at $100. Nobody sat down and decided to spend $240 a year on AI. It accumulated one reasonable-seeming signup at a time. One piece of local hardware eliminates all of it, permanently, for every person in the practice.

The hardware also has a 4–6 year useful life and does everything else a laptop or desktop does. It's not a pure AI line item. Really it's infrastructure.