All benchmark runsServing performance / Model endpoint

GLM-5.3 Flash

stock · Warm-engine unique-prefix pilot; about 4K input, client C1

completedPerformanceActual inference

glm53-flash-stock-serving-4k-c1-pilot-119f8a2564a2

Observed outcomes

6 / 8 observed passes

6 / 8 verified primary passes. Every declared case stays in the denominator; unverified attribution does not mean an observed task failed.

2fail

6pass

End-to-end completion throughput

21.183 tok/s

Server-reported completion tokens divided by the entire measured batch duration, including prefill, waiting and failed requests. Token accounting is stated in the serving cell.

8 samples

First generated output, median

2,563.189 ms

First nonempty reasoning or content delta; this may precede any visible answer. Includes every response with that timing; missing timings are not zero.

8 samples

First final content, median

3,322.551 ms

First nonempty final-content delta; arrival does not establish correctness or usefulness. Includes every response with that timing; missing timings are not zero.

8 samples

Request completion, median

4,155.302 ms

Full request duration, including reasoning; includes failures with recorded duration. Includes every response with that timing; missing timings are not zero.

8 samples

Request completion, p95

6,909.581 ms

Linear-interpolated p95 over all recorded request durations, including failures. Small samples give an unstable tail estimate.

8 samples

Stream chunk interval, median

87.761 ms

Observed SSE intervals reported by the serving client. Speculative decoding can emit several tokens per chunk; this is not per-token latency.

198 samples

Measured batch duration

37.153 s

Client measurement interval, excluding process setup. This is the denominator for aggregate completion throughput.

1 samples

Limitations

Prompt-only development workloads; these do not exercise native agent loops or establish held-out quality.
Input-size labels are targets. Actual server token usage differs across tokenizers and does not establish a maximum usable context window.
Client concurrency limits and configured server slots are not observed overlap or maximum sustainable capacity.
Observed overlap, when available, counts client request spans; it is not a count of active GPU sequences.
No throughput ranking or comparison delta is authorized by this record; matching conditions have not been established.
Transport completion and independently graded answer correctness are separate. Missing attempts and grades stay in the denominator.
First final-content arrival is a channel timing, not proof that the content is useful or correct.
Final-content timings exclude responses flagged for thinking tags in that channel, or lacking that observation; separate reasoning duration is not inferred.
Eight requests in one run are a pilot, not a throughput ceiling or a stable tail-latency estimate.
Completion-token usage combines reasoning and final output; the server did not report separate token counts.
Actual server input usage is used. The serving client reported different tokenizer counts.
The peak-overlap estimate is omitted because observer v3 lacks request start/end timestamps.
Correctness is taken from the separate frozen-evidence-fields-v1 receipt; original grading-script source identity was not recorded.
The schema task is a textual instruction in this workload; constrained response-format transport was not exercised.

Case results

Observed outcome and primary-model attribution are separate.

request-01Synthetic request 01pass

Primary outcome: pass
Failure stage: Not recorded
Duration: 3,923.377 ms

request-02Synthetic request 02fail

Primary outcome: fail
Failure stage: answer_checks
Duration: 6,931.26 ms

request-03Synthetic request 03pass

Primary outcome: pass
Failure stage: Not recorded
Duration: 3,801.035 ms

request-04Synthetic request 04pass

Primary outcome: pass
Failure stage: Not recorded
Duration: 4,393.57 ms

request-05Synthetic request 05fail

Primary outcome: fail
Failure stage: answer_checks
Duration: 4,387.226 ms

request-06Synthetic request 06pass

Primary outcome: pass
Failure stage: Not recorded
Duration: 3,203.412 ms

request-07Synthetic request 07pass

Primary outcome: pass
Failure stage: Not recorded
Duration: 3,635.69 ms

request-08Synthetic request 08pass

Primary outcome: pass
Failure stage: Not recorded
Duration: 6,869.321 ms

Deployment and test conditions

Configured capacity is distinct from successfully tested context or concurrency.

Model: GLM-5.3 Flash
Checkpoint revision: eb9eb208eb0d988989d07a6a12d0fdeb5f52574a
Variant: stock
Quantization: Native FP8 weights; FP8 E4M3 KV cache
Hardware: Four NVIDIA DGX Spark systems, GB10, 128 GB unified memory per system
Node count: 4
Runtime: SGLang
Runtime version: 0.0.0.dev1+gd6ab04bdf1
Configured context: 262144
Configured request slots: 8
Profile: Warm-engine unique-prefix pilot; about 4K input, client C1
Weight identity: Verified; see evidence
Suite version: serving-performance-v1
Started: Sep 20, 2026, 2:31 PM UTC
Finished: Sep 20, 2026, 2:32 PM UTC
Input-size target: 4096
Observed input-token range: 3704–3735
Client concurrency limit: 1
Observed peak request overlap: Not reported
Cache condition: Warm engine, unique prefixes
Transport completions: 8 / 8
Completion-token accounting: Reasoning and final content combined
Stream intervals: SSE chunks, not individual tokens
Timing observer SHA-256: 8a29437cb3e4b806f4be78454483721e9c31b09a7063f55bd9ec50081ffc3bef
Dataset SHA-256: 53ddfba9e794833f0ddb92f9785dc1c012c566ef6d438df0b4dd153ca074860f
Correctness scorer: frozen-evidence-fields-v1
Peak-overlap basis: Not recorded; serving-client peak estimates are excluded
Runtime image digest: sha256:f5de20f9b4a806a94012d405f5cb8734c129eb69a7b3875d5362210527bad575
Weight verification: All checkpoint files rehashed against the pinned artifact on all four replicas
Serving layout: TP4, four hosts; prefill chunk 2048; scheduler cap 8; static memory fraction 0.80
Observed KV token pool: 1204416 tokens total; not eight full-length sessions
Parsers: glm45 reasoning; glm47 tools
Speculation: Adaptive NEXTN: launch 5 steps/6 draft tokens; observed 3/4 at readiness
Thinking request: High reasoning; thinking enabled
Request settings: Temperature 0; seed 20260920; high reasoning; completion cap 4096 tokens; streaming with server usage
Input preparation: Identical frozen UTF-8 documents across tokenizers; approximate 4K target;8 independent development tasks
Cache state: Unique document prefixes first used on an already warm engine; no warmup requests or cache flush
Measurement path: CPU-only client over local network HTTP; no SSH tunnel in the measured inference path
Available scorer identity: Named scorer version; original grading-script source hash was not recorded

Provenance

Suite SHA-256: 7604be858e8b928dd7c2acb188dd25ec8b46bf26d544872220e18c20344db1d6
Evaluator SHA-256: 1f2d6733c17c647cba8599f8b560fdb2529b2fc9e710f8dd6ad71c68d606bf40
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.