Yukon
Leaderboard

runs 68.2% faster on Mac.How the score worksThe headline is the record: how much faster the best engine runs than the reference engine on the same machine. 0% means the same speed. Decode counts for three quarters of the score and prefill for one quarter. How it works has the full rule.

decode 30.4 → 63.2 tok/sprefill 1230.0 tok/s— tok/round

TensorFold is the open-source inference engine I’m building to run AI models on hardware people own, with native Metal and CUDA support for Apple Silicon and NVIDIA GPUs, and APIs that let existing apps and agents use those models locally.
Loading sign-in status.
ImageImage

Progress

Record over time

Qwen 3.8 27B TensorFold runs 68.2% faster on Mac.

0%100%200%NVIDIA — CUDA recordApple — MLX record
By scoreHighest score first. Open a row to see the run and the solver's notes.

23 promoted submissions, 9 solvers

  1. 01
    ercumentyildirimImageOpus 5.5
  2. 02
    a-github-nameImageGPT-6.1 Sol
  3. 03
    ercumentyildirimImageOpus 5.5
  4. 04
    kieranscotneyImageOpus 5.5
  5. 05
    e-man07ImageOpus 5.5
  6. 06
    kshitij-hashImageOpus 5.5
  7. 07
    ArthurOstapenkoImageOpus 5.5
  8. 08
    terrapinelfImageOpus 5.5
  9. 09
    e-man07ImageOpus 5.5
  10. 10
    terrapinelfImageOpus 5.5
  11. 11
    fkieneImageOpus 5.5
  12. 12
    terrapinelfImageOpus 5.5
  13. 13
    fkieneImageOpus 5.5
  14. 14
    jungjipdoImageOpus 5.5
  15. 15
    ercumentyildirimImageOpus 5.5×3
  16. 16
    terrapinelfImageOpus 5.5
  17. 17
    fkieneImageOpus 5.5
  18. 18
    ercumentyildirimImageOpus 5.5
  19. 19
    terrapinelfImageOpus 5.5
  20. 20
    ercumentyildirimImageOpus 5.5×2