Model artifact
Exact model name, revision, parameterization, quantization or precision, tokenizer, and serving configuration.
Performance claims
need a reproduction path.
SPARK does not publish universal latency, throughput, GPU-efficiency, or cost-reduction numbers without the model, hardware, workload, and measurement method needed to interpret them.
Review the protocol ↓NO INDEPENDENTLY VERIFIED PRODUCTION BENCHMARK PUBLISHED
Workload-specific reports may be produced during evaluation. A result moves to this public page only after disclosure and publication rights are approved.
MINIMUM DISCLOSURESix parts of a
Six parts of a
reviewable result.
Any future benchmark table will link to a test record containing these fields. A chart without them is treated as illustrative UI, not proof.
Hardware
Accelerator model and count, CPU and memory profile, interconnect, region, and tenancy.
Traffic
Input/output token distributions, context length, concurrency, request rate, warm-up, cache state, and duration.
Metrics
Time to first token, inter-token latency, end-to-end latency by percentile, throughput, errors, and cost basis.
Quality
Task-specific dataset, scoring method, evaluator version, human-review process, and confidence interval.
Reproduction
Date, code or configuration digest, exclusions, raw-output retention, and reviewer approval.
REPORT TEMPLATECompare against
Compare against
the customer baseline.
FIELDBASELINESPARK CANDIDATESTATUS
Quality scoreRequiredRequiredNot published
P50 / P95 latencyRequiredRequiredNot published
Output tokens / secondRequiredRequiredNot published
Error and timeout rateRequiredRequiredNot published
Cost basisRequiredRequiredNot published