Shareable preset

Private local lab

A capacity-first one-person setup for local-only coding, vision, and documents.

Your three planning routes

Start with the first one and argue with the other two.

Coding · Documents · Vision
Local only
Performance
Under $7,500
Heavy use

About the numbers: "runs locally" estimates how much of your chosen work the local machine takes on. Model quality is a separate question, and the benchmarks page covers it.

Tokens or towers

Cumulative cost, 36 months

The all-cloud baseline shows estimated API spend without buying new hardware.

View as table

Cumulative cost at six-month intervals
Month
0
6
12
18
24
30
36

Recommended

The sensible default

The best balance of capability, privacy, and money for the constraints you chose.

Hardware: ~$3,999
API / month: $0
Runs locally: 100%
Breaks even: —

MachineM3 Ultra Mac Studio96 GB unified · 819 GB/s

A quiet, turnkey complete system at list price for capacity-first MLX or llama.cpp serving.

Apple Mac Studio newsroom ↗

Model routes

DeepSeek V4 Flash 0731 (experimental Q2)local

coding + document work

~98 tok/s · bandwidth estimate

source ↗

Llama 4 Scoutlocal

document work + vision

~49 tok/s · bandwidth estimate

source ↗

Why it fits

The large memory pool holds private models that outgrow consumer GPUs.
Every selected workload has a local route; estimated API spend is exactly zero.

Reality check

Large DeepSeek quantizations on Spark are experimental and need hands-on serving work.
Treat hardware prices as planning allowances and check a live listing first.
Token speeds come from memory-bandwidth math at batch size 1. Measure before you rely on them.

Lowest total cost

The pay-as-you-go route

The lowest plausible total cost, with elastic APIs doing more of the work.

Hardware: ~$1,399
API / month: $0
Runs locally: 100%
Breaks even: —

MachineM4 Pro Mac mini24 GB unified · 273 GB/s

A quiet, turnkey complete system at list price for MLX or llama.cpp local inference.

Apple Mac mini newsroom ↗

Model routes

Gemma 4 31Blocal

document work + vision

~7 tok/s · bandwidth estimate

source ↗

Qwen3.6 27Blocal

coding + document work

~8 tok/s · bandwidth estimate

source ↗

Why it fits

24 GB is a practical entry point for useful quantized local models.
Every selected workload has a local route; estimated API spend is exactly zero.

Reality check

Treat hardware prices as planning allowances and check a live listing first.
Token speeds come from memory-bandwidth math at batch size 1. Measure before you rely on them.

Maximum privacy

Keep the work close

The most local coverage your hardware budget can reasonably support.

Hardware: ~$3,999
API / month: $0
Runs locally: 100%
Breaks even: —

MachineM3 Ultra Mac Studio96 GB unified · 819 GB/s

A quiet, turnkey complete system at list price for capacity-first MLX or llama.cpp serving.

Apple Mac Studio newsroom ↗

Model routes

DeepSeek V4 Flash 0731 (experimental Q2)local

coding + document work

~98 tok/s · bandwidth estimate

source ↗

Llama 4 Scoutlocal

document work + vision

~49 tok/s · bandwidth estimate

source ↗

Why it fits

The large memory pool holds private models that outgrow consumer GPUs.
Every selected workload has a local route; estimated API spend is exactly zero.

Reality check

Large DeepSeek quantizations on Spark are experimental and need hands-on serving work.
Treat hardware prices as planning allowances and check a live listing first.
Token speeds come from memory-bandwidth math at batch size 1. Measure before you rely on them.

Loading preset

Cumulative cost at six-month intervals

Month