> ## Documentation Index
> Fetch the complete documentation index at: https://docs.infercrane.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Benchmark evidence

> Inspect InferCrane performance claims with their exact model, runtime, hardware, workload, receipts, and limits.

# Claims with receipts

InferCrane binds every benchmark to an exact model, runtime, accelerator, workload, and date. A
measured win is not generalized to another provider, GPU, context length, concurrency level, or
production network path.

## Qwen3.8-27B on one H200

| Field | Value |
| - | - |
| Date | 2026-09-23 |
| Model | `Qwen/Qwen3.8-27B-FP8@017b9c7af6b5689d5dd426a76e0bc077eb5ca20a` |
| Runtime | SGLang 0.5.20 |
| Accelerator | 1× NVIDIA H200 |
| Workload | 4,096 input tokens, 512 output tokens, concurrency 12 |
| Selected recipe | FP8 weights and KV cache, bounded CUDA graphs, native NEXTN/MTP |

| Measurement | Control | Selected | Change |
| - | -: | -: | -: |
| Aggregate output throughput | 810.4 tok/s | 1,393.6 tok/s | 1.72× |
| Per-request p50 output speed | 76.1 tok/s | 141.0 tok/s | 1.85× |
| p50 TTFT | 1,591 ms | 739 ms | 53.6% lower |
| p95 inter-token latency | 14.02 ms | 8.67 ms | 38.1% lower |
| Successful requests | 24/24 | 24/24 | equal |

The paired run used the same model, image, workload, hardware, and output checks. Deterministic
outputs matched. Streaming and buffered usage, finish reasons, structured output, forced tools,
malformed-request handling, and prompt-token accounting passed.

* [Read the decision record](/testing/qwen38-openrouter-launch-decision-2026-09-23)
* [Inspect the machine-readable receipt](https://github.com/infercrane/infercrane/blob/main/docs/testing/evidence/qwen38-public-decode-saturation-modal-2026-09-23T044530Z.json)
* [Inspect raw measurements](https://github.com/infercrane/infercrane/blob/main/docs/testing/evidence/qwen38-public-decode-saturation-modal-2026-09-23T044530Z-raw.json)

<Warning>
  This is a controlled Modal H200 qualification, not an OpenRouter leaderboard or public-Internet
  result. The launch context boundary stayed at 32K because the near-262K run missed its
  pre-registered p95 TTFT limit. Hardware and workload changes require new evidence.
</Warning>

## Rejections are part of the result

The campaign preserved candidates that sounded attractive but lost under the target workload:

* FlashInfer GDN paths preserved correctness but regressed throughput and cold readiness.
* DFlash2 won some lower-concurrency screening lanes but did not replace native MTP at the selected
  saturation boundary.
* The near-262K context run completed correctly but missed its latency SLO.

InferCrane distinguishes modeled proposals, screening measurements, qualification evidence, and
public repeated evidence. Only comparable measurements can be ranked. Missing identity, quality,
cost, or workload evidence produces `WAIT` or `REJECT`; it is never converted into a release win.

Read [benchmarking](/features/benchmarking), [optimization](/features/optimization), and
[compatibility and qualification](/compatibility) for the evidence rules.