The open-weight models the radar often talks about can be downloaded and run on your own computer, free of charge and without sending a byte out. The two tools that made this accessible are Ollama (from the terminal: one command to download a model, one to talk to it) and LM Studio (an application with an interface: catalogue, downloads, chat, all clickable). Both free.

The right question first: why?

Three reasons hold up, the others usually don’t. Data that cannot leave: with a local model the text never leaves the machine, end of the cloud-and-policy discussion. Volume: if a task grinds through millions of tokens a day, even the cheapest API adds up, and a local model amortises the hardware. Understanding: running a model at home teaches you more about the nature of these systems than ten articles. If instead you are only after maximum quality, the honest answer is that frontier models do not run on a laptop: local is chosen for the constraints, and quality is bought in model size and therefore in hardware.

What it really takes

The variable is model size, and memory is the bottleneck. The figures declared by the projects: Ollama lists 8 GB of RAM for 7-billion-parameter models, 16 for 13B, 32 for 33B (official repo, checked on 19/07/2026); LM Studio recommends 16 GB of RAM and at least 4 GB of GPU (official requirements, same date). In practice: a recent laptop with 16 GB runs the small models well (up to ~14B quantised), and these are surprisingly capable on well-bounded tasks; the big models (70B and up) want workstations with 48-64 GB or dedicated GPUs, and there the maths must be done against API costs.

How they combine with the rest

The interesting part is that local tools speak the same language as the APIs: LM Studio and Ollama expose a compatible endpoint, so an agent like OpenCode can use the local model as its engine, and the method of your own cases stays identical: same test bench, local model in place of the remote one, and you see with your own eyes whether it is enough for your task.

The limits, said plainly

Quality per gigabyte is the real tax: a model that fits in 16 GB is not a frontier model, and on open-ended tasks the difference shows. Speed on pure CPU is modest; a fluid experience wants a GPU or a Mac with unified memory. And maintenance is yours: versions, quantisations, updates.

Testing: descriptive sheet; hardware requirements checked against official sources on 19/07/2026, no operational procedure included. The full run on a machine with declared specs (as the editorial plan requires) remains to be done: this container is not the right machine for an honest tokens-per-second figure.