Most people build a local AI workstation around the fastest processor they can justify, then discover that the model they wanted does not fit in memory, or operates at a snails pace. Now the same box may be asked to make images too.

The verdict up front: prove the workflow on the computer you already own, then buy memory for the jobs you will actually run at the same time. A 24GB to 32GB GPU is a starting range for compressed models. A larger unified-memory pool gives you more room, but neither capacity guarantees your particular workflow. Traveling? See running a local LLM on a travel mini PC.

Name the model before you size the machine

“Qwen3.8” is 3 different jobs: a 27B dense model under Apache 2.0, roughly 180-billion-parameter Flash Next, and a 2.4T flagship beyond a normal home build. For a 24GB to 32GB GPU, check a quantized 27B build, meaning compressed weights, with room for context. Flash Next needs substantially more memory; runtime and compression matter.

Ollama defaults to 4k below 24 GiB of VRAM, 32k from 24 to below 48 GiB, and 256k at 48 GiB or more. Longer context needs more memory. Check ollama ps for context and CPU/GPU split.

Images are now a local job

Open image models can run on your own hardware without a hosted per-image charge. Electricity, hardware and setup still count. FLUX.2 klein 4B is Apache-licensed, and Black Forest Labs reports as little as 13GB of VRAM. Z-Image Turbo, also Apache 2.0, says it fits 16GB consumer cards. Krea 2, a roughly 12-billion-parameter June release, uses Krea’s own gated community license, so read the terms before using it for a paycheck.

Do not assume 2 models that fit separately fit together. Check context and image encoders, or unload one before switching. Capacity is also not speed: AMD’s own test on a 128GB Ryzen AI Max+ 395 took about 78 seconds per 1024×1024 FLUX.1-dev image at full precision after warm-up.

Tower or unified memory?

In a tower, the RTX 3090 and 4090 carry 24GB. The RTX 5090 raises that to 32GB, with NVIDIA specifying 575W of graphics power and a 1000W system. Seat the power connector carefully: Club386 reported its own card’s cable melting in June.

Unified-memory machines share RAM between CPU and GPU. The GMKtec EVO-X2 uses Ryzen AI Max+ 395. Its purchase page defaults to 64GB; select a 128GB option deliberately. Framework’s 192GB Desktop with Ryzen AI Max+ PRO 495 is a Linux-only pre-order, initially one batch shipping in November. DwarfStar 4 supports Strix Halo, but lists Qwen3.8 Flash Next for Mac and NVIDIA only. Check memory upgradeability before buying.

How I actually use it

At home, I run a 9B Qwen model locally. On the road, I use ChatGPT and Claude, with OpenRouter picking up the overflow. I have been away for months, and the releases since have changed the picture enough that I will be retesting and upgrading the home lab when I get back.

Who should skip most of this

If local AI is occasional, or every important file already lives in shared cloud tools, a hosted model is simpler. If maintaining runtimes, licenses and updates sounds like a second job, skip it. Local software needs updates too: Ollama patched a model-loading flaw this year. Start with the exact model you will run, size memory second and speed third.

Cited sources


More field guides

More Remote Office guides

All Remote Office guides