Model cards · No. 5 of 17

KAT-Coder v2.5

Q8_0RetiredRetired

83.0Local AI Index / 100

KAT-Coder v2.5 (Q8_0) ranks #5 of 17 on the diego.technology Local AI Index with 83.0/100: 17.55/20 on twenty graded tasks (one run), 7.55/10 on the hard half, at 51.4 tokens/s on a 128 GB AMD Strix Halo machine.

The numbers

Measured, not quoted.

KAT-Coder v2.5 (Q8_0) ranks #5 of 17 on the diego.technology Local AI Index with 83.0/100: 17.55/20 on twenty graded tasks (one run), 7.55/10 on the hard half, at 51.4 tokens/s on a 128 GB AMD Strix Halo machine.

one run so far
17.55/ 20
one run so far
on the hard half
7.55/ 10
on the hard half
generation · 368 prompt
51.4tok/s
generation · 368 prompt
to load
15s
to load

Where the 83.0 points came from

  • Capability43.9 / 50
  • Hard problems15.1 / 20
  • Speed15.0 / 15
  • Consistency9.0 / 15

Strengths

Where it wins, where it slips.

Twenty graded tasks in six families. Brighter is better. Hatched means a run ran out of output budget before answering — unfinished, not wrong.

  • Code79%
  • Math96%
  • Data100%
  • Tools100%
  • Context100%
  • Reasoning67%
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20

Tasks 11–20 are the hard half. Hover a cell for the task.

Every task, as a list
  • 01 Python algorithm (interval merge)100%
  • 02 Python bug fix (date arithmetic)100%
  • 03 Compound-interest math100%
  • 04 Tiered tax calculation100%
  • 05 JavaScript ratio stress test100%
  • 06 SQL: best value per group100%
  • 07 Single tool call100%
  • 08 Bilingual email → JSON100%
  • 09 Long context (~18k tokens)100%
  • 10 Predict program output100%
  • 11 Payment schedule with prepayments83%
  • 12 Two-formula penalty comparison100%
  • 13 Expression parser, no eval()budget
  • 14 JS async retry with backoff100%
  • 15 Exact set cover100%
  • 16 SQL window functions100%
  • 17 Multi-file bug hunt71%
  • 18 Multi-hop long context100%
  • 19 Multi-step tool use100%
  • 20 Logic puzzle0%

The record

One full run.

A second run is what moves a setup from provisional to verified.

  1. 2026-09-2617.55 / 20 · hard 7.55 / 10