We took the top miner submission for Qwen3.8-27B and tested it against the pinned baseline using NVIDIA's AIPerf, an open-source benchmark for LLM inference. All 600 requests completed, and output throughput was 23 to 54% higher at every load we tested.

Peak output throughput
+54%
Requests completed
600 / 600
Lower GPU cost per token
19–35%

The setup

Model
Qwen3.8-27B
Serving stack
SGLang
Hardware
4× RTX 5090
Benchmark
NVIDIA AIPerf
Requests
8K-token input, 2K-token output
Method
Paired test against the pinned baseline

Paired means the patched engine and the baseline ran the same requests on the same machine, so the only thing that changed between the two runs was the miner's patch.

The results

Output throughput vs. baselineQwen3.8-27B on SGLang, 4× RTX 5090. Same requests, paired runs.
Output throughput vs. baseline
Concurrent requestsThroughput gain
4 concurrent+54%
8 concurrent+50%
16 concurrent+40%
32 concurrent+23%

The gain is largest at low concurrency and narrows as the GPUs get busier. A likely reason: with only a few requests in flight the GPUs have more idle capacity, and that is where better kernels and scheduling recover the most. Even at 32 concurrent requests the patched engine still served almost a quarter more tokens on the same hardware.

Why it matters

Every token costs money. If the same four GPUs serve 23 to 54% more output, the GPU cost behind each token falls by roughly 19 to 35%, with no new hardware and no change to the model.

GPU cost per token, reduction vs. baselineDerived from the throughput gains above: 1 − 1 / (1 + gain).
GPU cost per token, reduction vs. baseline
Concurrent requestsCost reduction
4 concurrent−35%
8 concurrent−33%
16 concurrent−29%
32 concurrent−19%

That is the point of running this on Subnet 10. Miners compete to make the software that serves a model faster, and only changes that hold up against a pinned baseline get kept. The result here was not produced by one team over a quarter. It came out of an open competition, and the next round starts from it.

Every round and submission is public on the dashboard. The scoring rules are in the docs.