Methodology

localLLMrig answers one question: does this model fit on this machine, and roughly how fast? The answer is an estimate. Where we opened a page, the URL sits next to the figure. Where we did not, the source field stays empty.

Weights

Weight size ≈ parameters × bytes per parameter. We use 0.56 (Q4), 0.69 (Q5), 1.07 (Q8) and 2 (FP16). That is close to GGUF Q4_K_M, Q5_K_M, Q8_0 and FP16, plus a little container overhead. It is not a measured file size.

The parameter count itself is not in the Hugging Face config. That field is therefore always an estimate, even when the layers and KV heads come from a config.

KV cache and context

KV cache in FP16: 2 × layers × kv heads × head_dim × 2 bytes × tokens. For Gemma 2 we treat half the layers as a sliding window, and that is an approximation. On top of that, 0.8 GB of runtime overhead.

What counts as a fit

On a discrete GPU we subtract 1.2 GB per card. On a Mac we leave room for the system: 6 GB at 16 GB, 8 GB at 32 GB, 10 GB at 64 GB, 12 GB above that. If the total fits, it is a full fit.

If it does not fit in the GPU but does fit in GPU plus system RAM (minus 8 GB for the OS), it is CPU offload. That option stays out of the top 3 while a full fit exists. It can show up as a footnote, because it is slow.

Tokens per second

We assume decode is memory-bandwidth bound: tokens/s ≈ 1 / (gigabytes / effective bandwidth + 0.0025 s). Effective bandwidth is the spec sheet times a utilisation: NVIDIA 0.55, Apple 0.48, AMD 0.4. With two GPUs we also multiply by 0.8, because a layer split does not add bandwidth.

That utilisation is an assumption, not a measurement. Small models on a very fast card are often compute-bound, so the estimate can be optimistic there. There is no benchmark source, because we don't have one.

Prices

Every price is an indicative EU amount incl. VAT, dated 2026-10-14. The source field is empty. We did not store a bol.com, Coolblue or Apple page as a price source. Used prices move the most.

How the top 3 is chosen

Full fits first. Inside the budget: the fastest, the best tokens-per-euro, and another platform if there is one. If spots remain, we add a faster machine above the budget. If nothing fits inside the budget, we show the nearest full fits and say they sit above it.

Machine-readable

Agents can read /llms.txt and /llms-full.txt, plus JSON at /api/models, /api/hardware and /api/recommend. Those payloads say outright that memory and tokens/s are estimates, and that price and benchmark sources stay empty.

Longer English guides live at /guides.

What is not in here

No electricity cost, no decibel rating, no 'this shop has it in stock'. No Q3 or Q6. No multi-user batching. The Mac line follows the pages we opened in October 2026 (M5/M6) plus M4 variants from the 2025 Studio, because those are still the used-market question.