Deploy Qwen on an H100.

Qwen was missing from the site even though the public H100 work used it. This page names the checkpoint that was actually timed.

What was measured

Qwen3.8-27B, one H100, one request at a time.

• In the 23-way comparison, that H100 placed 3rd. Median output speed on the published board was 241 tok/s. Against a typical provider in that test it was 2.6–3.9×, depending on the prompt. Full method: benchmark.
• In the earlier tuning run, the same class of GPU went from 50.8 tok/s at 16-bit to 213 tok/s with FP8 weights and DFlash2 drafting 7 tokens. Method: 213 tok/s.

Those are not “Qwen3” in the abstract. A different size, a different quant, or a batch of 16 requests will not match. The tuning post records 130 tok/s per request at 16-wide, and 1,817 tok/s for the GPU as a whole, on that one H100.

How denv deploys it

A recipe pins the FP8 weights revision, the drafter revision, the draft length, and the vLLM image digest. The tuning post names vLLM 0.32.0 for that night. A setup counts only after it has come up healthy on that GPU type. H100 sites are Lambda, Nebius, EC2, or your server. The 50.8 start was a Nebius account.

You pay the GPU provider. denv does not sell tokens. Health checks and failover are the same as for any other recipe. SGLang is available when a recipe uses it. The published Qwen numbers were vLLM.

Other Qwen checkpoints can be a custom recipe. They do not inherit these speeds. The product is early.