1.28x faster median latency on H100, vLLM 0.25.1, concurrency 8, recorded Jul 25, 2026.
1 measured package. Results apply to the recorded conditions.
Packages are no longer sold. For hosted inference, see Model APIs.
| Package | GPU | Engine | Optimization | Verified |
|---|---|---|---|---|
| H100, Channelwise FP8, measured selective-layer recipeqwen3-6-27b-fp8cd-v3-mlponly-h100-vllm | GPUH100 | EnginevLLM 0.25.1 | OptimizationChannelwise FP8, measured selective-layer recipe | VerifiedJul 25, 2026 |
Selective channelwise FP8 leaves the sensitive paths at higher precision while converting the remaining eligible layers.
Published Jul 25, 2026. Proof verified Jul 25, 2026.
chat-128
vLLM 0.25.1. Concurrency 8. Proof verified Jul 25, 2026.
Median latency measured 2,857 ms at baseline and 2,215 ms optimized.
gsm8k 99.87% recovery, passed.
Published package measurements for Qwen3.6 27B, grouped from the public RunInfra catalog.
Measured by RunInfra on H100 with vLLM 0.25.1; proof dates are listed with each package.
We report package measurements here, and each package remains subject to its listed license.
Citation: RunInfra (2026). Published package measurements and signed kit records. https://runinfra.ai/model.
Other GPU measurements for this model not published.
Other engine measurements for this model not published.
Measurements for context lengths not listed in these packages not published.
No. Package sales are closed. For hosted inference, see Model APIs at runinfra.ai/inference-api.
H100, Channelwise FP8, measured selective-layer recipe is measured on H100 with vLLM 0.25.1. Proof verified Jul 25, 2026.
gsm8k 99.87% recovery, passed. Proof verified Jul 25, 2026.
Use a workspace API key and pay for input, cached input, and output tokens.
View Model APIs