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.
- Confidence is computed from the sources, never asserted —
verified,single-source,disputedorestimated. - Every measurement is dated and sortable, because the same hardware gets faster as backends improve.
- Anyone can add to it or argue with it — submit a benchmark, correct a number, or say what is wrong with the whole approach.
- 220measured runs
- 300citations
- 20sources
- 132contributors
- 47verified
- 163single-source
- 10disputed
- 90hardware
- 63models
- 1914estimated
What should I buy?
Uses measured runs where they exist; unmeasured picks are marked as an estimate. Full ranking on the advisor.
Ranked recommendations
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.
| Hardware | Model | Quant | Decode tok/s | Prefill tok/s | Measured | Evidence | Source |
|---|---|---|---|---|---|---|---|
 AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo) | Qwen3.8 27B | Q4_K_M | 9.8 | — | 2026-08-30 | single-source: One credible source | nabe2030 |
 RTX 5070 Ti + RTX 3070 Ti (mixed dual-GPU) | Qwen3.8 27B | Q4_K_M | 29.7 | 344.2 | 2026-08-24 | single-source: One credible source | h-nabata |
| AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo) | Qwen3.8 27B | Q4_K_M | 12.3 | — | 2026-08-22 | single-source: One credible source | julianmb |
| AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo) | Qwen3.8 27B | ROCmFP4_FAST | 34.8 | — | 2026-08-22 | single-source: One credible source | julianmb |
| AMD Ryzen AI Max+ 395 (128GB unified, Strix Halo) | Qwen3.8 27B | ROCmFP4_FAST | 14 | — | 2026-08-22 | single-source: One credible source | julianmb |
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.
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.