$ cat models/glm-5.3
GLM 5.3
Z.AI
Check that an endpoint selling GLM 5.3 really serves it. A test compares the endpoint's answers with answers collected from Z.AI, through OpenRouter pinned to Z.AI.
- Official Provider
- Z.AI, through OpenRouter pinned to Z.AI
- Reference Answers
- 200
- Declared Precision
- fp8
- Official Release
- fp8 e4m3, 128x128 blocks
- Checks
- Identity · Competence · Refusals (9 probes)
- Reference Data Updated
- 2026-10-16
How Well a Test Works
Size
Requests
Genuine Confirmed
Wrongly Flagged
Substitutes Caught
Standard
80
99.5%
0.5%
99.9% · hardest GPT-6 Astra
Thorough
160
99.5%
0.5%
100.0%
Simulated from the reference answers. A substitute that isn't caught ends inconclusive: it is never confirmed as this model.