GLM-5.3-Flash 524K xgrammar qualification evidence¶
These artifacts support the dated finding. They are sanitized, portable evidence from the dual RTX PRO 6000 Max-Q TP=2/DCP=2 qualification. Private router URLs, host identities, GPU UUIDs, credentials, and real-client traces are intentionally excluded.
The derivative publication summary maps every compact public claim back to this finding and the retained artifacts.
Feasibility¶
feasibility-input.jsonrecords sourced and explicit-unknown fit inputs.feasibility-result.mdrecords the benchmark-survivor decision and rejected assumptions.
Matched no-speculation control¶
nospec-preflight-text-tools-250k.jsonnospec-vision-ocr.jsonnospec-c1-ctx4096-r5.jsonnospec-c1-ctx240000-r5.jsonnospec-c2-ctx250000-max8192-r2.json
Corrected DFlash2 K5 candidate¶
dflash2-preflight-text-tools-250k.jsondflash2-vision-ocr.jsondflash2-quality-all-high-r3.jsondflash2-c1-ctx4096-r5.jsondflash2-c1-ctx240000-r5.jsondflash2-c1-ctx240000-r5-repeat2.jsondflash2-c2-ctx250000-max8192-r2.json
The capacity artifacts use capacity-v4-reasoning. Decode values include
reasoning completion tokens where the API reports them and are not raw
server-side token timestamps. The C2 artifacts prove bounded concurrent
completion, not a clean output-rate comparison, because response lengths and
reasoning behavior varied.