What do you want to run in the EU?

Five things people run at home, graded A / B / C / — for every listing from its specs alone. Pick one and a budget.

The grades

• Chat assistant: an assistant you talk to in your own words, running Llama, Qwen, Gemma or Mistral weights locally. A: a 14B-class model at reading speed or better. B: a 7–8B model at a usable pace. C: a 3B model, slowly.
• Coding assistant: completion and chat about code in an editor, where a bigger model and a long context matter. A: a 24B-class coder with room for a long context, fast. B: a 14B coder at a usable pace. C: a 7B coder, slowly.
• Image generation: Stable Diffusion and FLUX: compute-bound work that wants a GPU or a big unified-memory machine. A: FLUX at a working pace. B: SDXL, with patience for FLUX. C: Stable Diffusion 1.5, slowly.
• Voice transcription: Whisper and its descendants: meetings, dictation and voice notes transcribed on your own machine. A: large-v3-turbo faster than real time. B: the medium model at about real time. C: the small model, slower than real time.
• Vision (cameras, home automation): object detection on security cameras and a small vision-language model to describe what it sees, running all day. A: several cameras plus a 7B vision model, under 150 W all day. B: detection on a few cameras and a small vision model. C: detection on one or two cameras, on the CPU.

The formulas and thresholds are on how we score.