Can your machine run it?

How this checker works

How does the math work?

We add up what the model actually needs — the measured quantized weight file, the KV cache for your context length (computed from the model's real attention config), and runtime overhead — and compare it against your usable memory pool: VRAM for discrete GPUs, the GPU-addressable slice of unified memory for Macs and mini-PCs. Under ~70% of the pool grades A+, under 90% grades B, up to 100% is C (tight), and past the pool it either spills to system RAM via CPU offload or won't fit. The full formulas and thresholds are published on our methodology page.

Why a speed tier instead of a tokens-per-second number?

Because a precise number would be fake precision. Decode speed depends on runtime version, drivers, batch size, and thermals — so we compute a bandwidth-based estimate internally and publish only the tier it lands in: interactive (chat feels instant), usable (fine for most work), or painful. When real-world measurements accumulate, tiers get calibrated against them.

What does the beta badge mean?

It means we can tell you whether the model fits, but the speed tier isn't calibrated yet for that regime — AMD and Intel GPUs (ROCm/Vulkan), the new mini-PC class (DGX Spark, Strix Halo), MoE models, and CPU offload all carry wider error bars. We show the beta label instead of pretending. Submitting your real tokens-per-second helps us calibrate.

Why do you show used prices?

Because the used market is where local-LLM hardware actually gets bought — a used RTX 3090 is the canonical budget build. Every price carries an as-of date and a source; we'd rather show a dated real number than a fresh made-up one.

How do I submit my real results?

Run a result, then use the submission form under it: hardware, model, quant, and runtime are pre-filled — you add your measured decode tokens-per-second. Submissions are reviewed before they influence any tier (that's our anti-poisoning policy), and once enough accumulate we publish our median error.