Local AI Arena

Which machine runs the model you want, how fast, and what it costs. Every figure is a real measurement with a named source and a date.

What should I buy?

Uses measured runs where they exist; unmeasured picks are marked as an estimate. Full ranking on the advisor.

Form factor

Leave all unchecked for no constraint.

Your ranked recommendations will appear here after you press Recommend.

Browse the data How the numbers are judged About this project

Most recent measurements

Decode is token generation — the speed you feel while an answer streams. Prefill is prompt processing — the wait before it starts.

HardwareModelQuantDecode tok/sPrefill tok/sMeasuredEvidenceSource
AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo)Qwen3.8 27BQ4_K_M12.32026-08-22single-source: One credible sourcejulianmb
AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo)Qwen3.8 27BROCmFP4_FAST34.82026-08-22single-source: One credible sourcejulianmb
AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo)Qwen3.8 27BROCmFP4_FAST142026-08-22single-source: One credible sourcejulianmb
NVIDIA GeForce RTX 3060 12GBLlama 2 7BQ4_075.62407.92026-08-21single-source: One credible sourcegrgsim
NVIDIA GeForce RTX 4090Qwen3.8 27BQ4_K_M47.42026-08-15verified: Two or more independent sources agreeing within 25%Spadav , snakeyjay63-png

See all 208 measurements

How the evidence accumulated

Every measured record, stacked by how well corroborated it is, cumulative by the month the benchmark was run. Estimates are excluded: they are generated rather than measured and carry no date of their own.

Cumulative measured benchmark records by monthA stacked bar chart running from 2023-11 to 2026-08, rising to 208 measured records, of which 47 are verified by two or more independent sources, 151 rest on a single source and 10 are disputed.0521041562082023-11: 8 measured — 6 verified, 2 single-source, 0 disputed2024-01: 12 measured — 6 verified, 6 single-source, 0 disputed2024-05: 13 measured — 6 verified, 7 single-source, 0 disputed2024-07: 14 measured — 6 verified, 8 single-source, 0 disputed2024-08: 15 measured — 6 verified, 9 single-source, 0 disputed2024-11: 18 measured — 9 verified, 9 single-source, 0 disputed2024-12: 19 measured — 9 verified, 10 single-source, 0 disputed2025-01: 24 measured — 10 verified, 13 single-source, 1 disputed2025-02: 28 measured — 10 verified, 17 single-source, 1 disputed2025-03: 39 measured — 14 verified, 24 single-source, 1 disputed2025-04: 43 measured — 14 verified, 28 single-source, 1 disputed2025-05: 49 measured — 18 verified, 30 single-source, 1 disputed2025-06: 52 measured — 19 verified, 31 single-source, 2 disputed2025-07: 57 measured — 21 verified, 34 single-source, 2 disputed2025-08: 84 measured — 31 verified, 47 single-source, 6 disputed2025-09: 100 measured — 33 verified, 60 single-source, 7 disputed2025-10: 126 measured — 38 verified, 81 single-source, 7 disputed2025-11: 130 measured — 40 verified, 83 single-source, 7 disputed2025-12: 139 measured — 41 verified, 91 single-source, 7 disputed2026-01: 141 measured — 42 verified, 92 single-source, 7 disputed2026-02: 152 measured — 44 verified, 101 single-source, 7 disputed2026-03: 171 measured — 44 verified, 118 single-source, 9 disputed2026-04: 179 measured — 45 verified, 125 single-source, 9 disputed2026-05: 186 measured — 45 verified, 132 single-source, 9 disputed2026-06: 196 measured — 46 verified, 140 single-source, 10 disputed2026-07: 203 measured — 46 verified, 147 single-source, 10 disputed2026-08: 208 measured — 47 verified, 151 single-source, 10 disputed2023-112026-08
47 verified 151 single-source 10 disputed

Read the shape with care. The taller months are partly a record of which sources have been mined rather than of when the world ran more benchmarks, and a run only enters here once someone finds and verifies it — so the right-hand edge understates recent months and will keep filling in.