Model cards · No. 1 of 17

gpt-oss-120b

MXFP4Current teamBuilder · daily driver

98.4Local AI Index / 100

Only ever dropped points on the expression parser — and still passed most of its hidden cases. Near-perfect at 4–5× Gemma’s speed.

The numbers

Measured, not quoted.

gpt-oss-120b (MXFP4) ranks #1 of 17 on the diego.technology Local AI Index with 98.4/100: 19.88/20 on twenty graded tasks (mean of 3 runs), 9.88/10 on the hard half, at 48.5 tokens/s on a 128 GB AMD Strix Halo machine.

mean of 3 runs
19.88/ 20
mean of 3 runs
on the hard half
9.88/ 10
on the hard half
generation · 248 prompt
48.5tok/s
generation · 248 prompt
to load
24s
to load

Where the 98.4 points came from

  • Capability49.7 / 50
  • Hard problems19.8 / 20
  • Speed14.9 / 15
  • Consistency14.1 / 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.

  • Code98%
  • Math100%
  • Data100%
  • Tools100%
  • Context100%
  • Reasoning100%
  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 prepayments100%
  • 12 Two-formula penalty comparison100%
  • 13 Expression parser, no eval()88%
  • 14 JS async retry with backoff100%
  • 15 Exact set cover100%
  • 16 SQL window functions100%
  • 17 Multi-file bug hunt100%
  • 18 Multi-hop long context100%
  • 19 Multi-step tool use100%
  • 20 Logic puzzle100%

As a coding agent

Six real repo tasks, four harnesses.

Hidden tests decide the score. Minutes are wall-clock for all six tasks.

Pi
6/ 6
4 min
OpenCode
6/ 6
6 min
dsh
6/ 6
6 min
Hermes
6/ 6
11 min

The record

3 full runs.

Repeat runs are what separate a daily driver from a lucky score.

  1. 2026-09-2719.96 / 20 · hard 9.96 / 10
  2. 2026-09-2819.72 / 20 · hard 9.72 / 10
  3. 2026-09-2919.96 / 20 · hard 9.96 / 10