Model cards · No. 4 of 17
GLM-5.3-Flash
UD-IQ3_XXSContenderHeavyweight contender
89.2Local AI Index / 100
Froze the GPU twice on a Vulkan build; perfect on ROCm. One run so far.
The numbers
Measured, not quoted.
GLM-5.3-Flash (UD-IQ3_XXS) ranks #4 of 17 on the diego.technology Local AI Index with 89.2/100: 20/20 on twenty graded tasks (one run), a perfect score, at 13.4 tokens/s on a 128 GB AMD Strix Halo machine.
- one run so far
- 20/ 20
- one run so far
- on the hard half
- 10/ 10
- on the hard half
- generation · 60 prompt
- 13.4tok/s
- generation · 60 prompt
- to load
- 45s
- to load
Where the 89.2 points came from
- Capability50.0 / 50
- Hard problems20.0 / 20
- Speed10.2 / 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.
- Code100%
- Math100%
- Data100%
- Tools100%
- Context100%
- Reasoning100%
- 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)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()100%
- 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%
The record
One full run.
A second run is what moves a setup from provisional to verified.
- 2026-09-3020 / 20 · hard 10 / 10