Model cards · No. 13 of 17
Gemma 4 26B-A4B QAT
UD-Q4_K_XLRetiredRetired
71.0Local AI Index / 100
The fastest model tested. All six misses were budget exhaustion.
The numbers
Measured, not quoted.
Gemma 4 26B-A4B QAT (UD-Q4_K_XL) ranks #13 of 17 on the diego.technology Local AI Index with 71.0/100: 14/20 on twenty graded tasks (one run), 6/10 on the hard half, at 57.5 tokens/s on a 128 GB AMD Strix Halo machine.
- one run so far
- 14/ 20
- one run so far
- on the hard half
- 6/ 10
- on the hard half
- generation · 635 prompt
- 57.5tok/s
- generation · 635 prompt
- to load
- 6s
- to load
Where the 71.0 points came from
- Capability35.0 / 50
- Hard problems12.0 / 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.
- Code67%
- Math100%
- Data100%
- Tools100%
- Context50%
- Reasoning0%
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 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)budget
- 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 outputbudget
- 11 Payment schedule with prepayments100%
- 12 Two-formula penalty comparison100%
- 13 Expression parser, no eval()budget
- 14 JS async retry with backoff100%
- 15 Exact set coverbudget
- 16 SQL window functions100%
- 17 Multi-file bug hunt100%
- 18 Multi-hop long contextbudget
- 19 Multi-step tool use100%
- 20 Logic puzzlebudget
The record
One full run.
A second run is what moves a setup from provisional to verified.
- 2026-09-2714 / 20 · hard 6 / 10