Local AI Index

• Score blends three groups: AI power 50%, experts 30%, users 20%.
• AI power comes from the specs: GPU memory (16–512 GB) and memory speed (64–1024 GB/s), each on a log scale. FP16 TFLOPS is counted once at least 3 in 4 machines have it; most makers do not publish it, so today it is shown only.
• Each review site is compared only with its own other ratings, so strict and generous sites count the same.
• 80 is the field average. Machines with few reviews are pulled toward 80. A hollow score marker means 0–2 reviews or buyer ratings. Shop ratings from the maker's own store are shown but not counted.
• Every machine gets a score from its specs. Only machines with at least one review or buyer rating get a rank; the others sit under the line.
• The rank number is the rank inside the tab you are on. All mixes every chip maker.
• Big-model fit (0–5) adds one point each for 48 GB+ of GPU memory (a 70B model fits at 4-bit), 96 GB+ (a 120B model fits), memory speed of 250 GB/s+, software (CUDA or MLX; ROCm or Vulkan only count half) and a fast link for clusters (10 GbE, Thunderbolt 5, USB4 v2 or QSFP).
• GPU memory is the most the maker's own tools let the GPU use without system changes. The note in each row says where the number comes from and what tuning allows, where known.
• Est. 70B speed is memory speed × 0.65 ÷ 42.5 GB, in tokens per second, for a 70B model at Q4_K_M. Measured speeds from reviews are in each row.
• Value is (score − 70) × (median price of all machines ÷ its price). Higher means more for your money. Its shade shows the rank among all machines.
• Vienna price is the lowest new price in an Austrian shop or on Geizhals.at. EU marks the maker's own EU shop and DE a German shop, when nothing Austrian turned up. Prices change daily.

Gaps: 0 Apple machines, 4 AMD machines, 3 NVIDIA machines and 1 Intel machine have no reviews or buyer ratings yet, and 12 machines have no price yet. They show a dash.

113 sites were used. Every number links to its source inside its row.