Tell us what you're running. We'll tell you which local models actually fit.
Every result below is checked against your memory, disk space, and what you're using it for — no guesswork about whether a download will actually run.
Your machine
How the fit — and the benchmarks — are calculated
If you have a dedicated GPU, we check the model against your VRAM first — that's what determines whether it runs at real-time speed. On Apple Silicon, RAM and GPU memory are the same pool, so we reserve about 4 GB for the system and use the rest. With no dedicated GPU, models run on your CPU using system RAM, which works but is considerably slower.
File sizes are real — read directly off each model's GGUF files on Hugging Face, not estimated. The "needs" figure adds a fixed overhead on top for the runtime and context (whichever is bigger: 1.5 GB, or 15% of the file). Actual usage still varies with context length and settings, so treat it as a planning number, not a guarantee. Vision-capable models also need a separate image-projector file that isn't counted in the size shown.
In CPU-only mode, the tokens-per-second number is a rough estimate, not a benchmark. Decoding on a CPU is mostly limited by memory bandwidth, not raw compute, so we estimate speed as roughly (your RAM bandwidth ÷ the model's file size), halved to account for real-world overhead. We guess bandwidth from the CPU name you enter — a workstation-class chip (Threadripper, EPYC, Xeon) or an Apple Pro/Max/Ultra tier both point to a wider memory bus than a typical consumer chip. Leave the field blank and we assume a generic dual-channel setup.
Benchmark scores come from each model's own Hugging Face model card, next to the area they measure. Scores marked † are cited in another model's own published comparison table rather than the source model's card — usually because the source card links out to a blog post instead of publishing numbers directly. Not every model has a published score for every area it's tagged for; we show what's actually published rather than estimate.
None of the runtimes below use a Neural Processing Unit (NPU) for LLM inference yet, including Apple's Neural Engine. NPUs accelerate other on-device AI tasks today, but the model formats here run on CPU or GPU.