About TokenAssemble

TokenAssemble answers one question: can your machine run it? Whether a local LLM fits on a given GPU, Mac, or mini-PC, how well it will feel to use, and what the best quant is — computed from measured file sizes and published specs, with every assumption stated.

The local-LLM world is full of confident wrong answers: calculators that ignore the KV cache, benchmark numbers quoted without their runtime version, “it runs” claims that mean 3 tokens per second. Our position is simple — an honest estimate beats a precise-looking guess. That's why speed is a tier instead of a number, why tight fits say “tight” instead of “yes”, why uncalibrated regimes carry a visible beta label, and why the answer is sometimes a plain “won't fit” with nothing to sell you.

Today we track 36 GPUs, 14 prebuilt systems (Apple Silicon and the unified-memory mini-PC class), and 40 models with their published quants — every row stamped with a source link and an as-of date (current snapshot: 2026-07-11). The formulas and thresholds behind every verdict are published in full on the methodology page.

The site is free. It earns money from clearly-disclosed affiliate links on hardware we already recommend on the merits — grades and tiers are computed from the data before any link is attached, and that ordering never reverses. AI tooling helps us polish prose; it never generates the facts.

Found a verdict that doesn't match your real-world experience? That's exactly the feedback the calibration loop needs — use the submission form under any result, or write to hello@tokenassemble.com.