What the LLM hardware recommender does
Pick a model (Llama 3.1, Qwen 2.5, Mistral, DeepSeek), set the minimum tokens/sec you'll tolerate, and this tool returns the cheapest Mac M-series or NVIDIA / AMD GPU that runs it smoothly — backed by 80+ measured benchmarks rather than spec-sheet guesses. It answers "what's the cheapest thing that runs this model fast enough for me?"
How to find your hardware
- Choose the model (and size) you want to run.
- Set your minimum acceptable speed in tokens per second.
- Get the cheapest Mac or GPU that clears that bar, with the measured throughput.
What decides whether a model runs
Two things: VRAM / unified memory — the model's weights (at your quantization) have to fit, or it won't load at all — and memory bandwidth, which sets how fast tokens come out once it fits. That's why an Apple M-series with lots of unified memory can run larger models than a GPU with less VRAM, even if the GPU is "faster" on paper. To decide whether to buy at all, run the self-host vs API calculator; the raw numbers live in the local LLM benchmark database.
