All benchmark runsServing performance / Model endpoint

GLM-5.3 Uncensored

uncensored · Timing-only C1; eight naturally stopping requests per input target; endpoint usage unknown

completedPerformanceActual inference

fullglm-frozen-4k-c1-f9f1db27fb75

Run JSON Case results CSV Model source

Observed outcomes

8 / 8 requests completed

Timing-only run. Answer correctness was not measured; no quality score is assigned.

8answers ungraded

End-to-end completion throughput

16.869 tok/s

Observed completion tokens, including reasoning, divided by the full measured batch duration. Includes waiting, prefill, generation and private LAN latency; not decode TPS.

8 samples

First generated output, median

2,781.899 ms

First nonempty reasoning or content payload observed by the client. This is not server-internal token timing.

8 samples

First final content, median

4,236.473 ms

First nonempty content payload among complete answers independently checked for reasoning tags throughout their captured text. Correctness and usefulness were not graded.

8 samples

Request completion, median

5,260.761 ms

Full client-observed request duration, including private LAN latency and any endpoint queueing.

8 samples

Measured batch duration

44.521 s

Client measurement interval excluding process setup; the throughput denominator.

1 samples

Limitations

These 16 timing requests are separate from the 192 native-agent quality attempts; they do not change either quality score.
Every timing request completed. Case outcomes mark answer checks as not run, not a transport failure or a zero accuracy score.
C1 only; no maximum context, concurrency or sustained decode-throughput claim.
Short naturally stopping responses and unverified competing endpoint traffic limit comparability.
Completion tokens include reasoning. First generated and first visible content are separate measurements.
Frozen observer version 4 checks only the first 256 content characters for reasoning tags; first-final-content metrics additionally use a complete saved-answer audit. Frozen receipts were not rewritten.
Checkpoint verification refers to the matching September 19 byte-hash and staging receipts; live tensors were not rehashed during this timing audit.
This frozen campaign does not evaluate subsequent application repairs.

Case results

Observed outcome and primary-model attribution are separate.

request-01Synthetic timing request 01Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 4,806.593 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-02Synthetic timing request 02Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 4,614.312 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-03Synthetic timing request 03Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 5,430.196 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-04Synthetic timing request 04Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 5,091.326 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-05Synthetic timing request 05Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 4,867.198 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-06Synthetic timing request 06Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 5,858.178 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-07Synthetic timing request 07Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 7,877.238 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

request-08Synthetic timing request 08Answer not graded

Primary outcome: unverified
Failure stage: answer_checks_not_run
Duration: 5,969.217 ms

Recorded mechanism checks

Transport completed
answer correctness was not graded.

Deployment and test conditions

Configured capacity is distinct from successfully tested context or concurrency.

Model: GLM-5.3 Uncensored
Checkpoint revision: 409ddf484d16cb7b13d390e8b30942890164d5af
Variant: uncensored
Quantization: NVFP4 weights; FP8 MLA KV cache
Hardware: NVIDIA DGX Spark GB10, 128 GB per node
Node count: 16
Runtime: vLLM with upstream structured-output repair
Runtime version: v19 + upstream PR44993
Configured context: 1048576
Configured request slots: 16
Profile: Timing-only C1; eight naturally stopping requests per input target; endpoint usage unknown
Weight identity: Verified; see evidence
Suite version: static-performance-evidence-1
Started: Sep 21, 2026, 10:42 PM UTC
Finished: Sep 21, 2026, 10:43 PM UTC
Input-size target: 4096
Observed input-token range: 3704–3735
Client concurrency limit: 1
Observed peak request overlap: 1
Cache condition: Cache state unreported
Transport completions: 8 / 8
Completion-token accounting: Reasoning and final content combined
Stream intervals: Not measured
Endpoint usage: Unknown; private LAN endpoint. AEVAmon was paused, but exclusive endpoint use was not verified.
Cache condition: Warm serving engine; no cache flush or verified prefix-cache state.
Request limit: One request at a time, eight requests per input target; 300-second request deadline.
Answer checks: Correctness not graded; timing evidence only. No quality score or ranking.
Timing observer: sglang-content-observation-4; complete captured answers independently checked for late reasoning tags.
First-final-content coverage: 8/8 complete answers without a <think> tag; 0 tagged, 0 truncated, 0 length mismatches.
Timing observer SHA-256: 2977430a164f9ef681d6cff4dd7e5826318f55d95d9ffe8740bce494dec505a9
Dataset SHA-256: 53ddfba9e794833f0ddb92f9785dc1c012c566ef6d438df0b4dd153ca074860f
Private timing receipt SHA-256: f9f1db27fb7514cacd6d0b33aefaf0a863e7cf8ee168c8632d1df1e808d67d62
Observed output-token range: 56–143
Checkpoint verification: 282 checkpoint shard hashes verified September 19 against the recorded upstream revision; matching source and staged-copy receipts retained.

Provenance

Suite SHA-256: 53ddfba9e794833f0ddb92f9785dc1c012c566ef6d438df0b4dd153ca074860f
Evaluator SHA-256: 2977430a164f9ef681d6cff4dd7e5826318f55d95d9ffe8740bce494dec505a9
Comparison conditions SHA-256: Comparability not established

No additional public evidence links were attached. The run JSON contains all published fields.

Public records exclude production conversations, private memories, network endpoints and hidden model reasoning.