Deploy Qwen on H200.

If you have an H200, usually at Nebius, you can ask for a Qwen recipe tested on that GPU. denv has not published a tokens-per-second result for Qwen on H200. The H100 results are H100 results. Copying 213 or 241 onto this page would be false.

DFlash2’s authors report their own acceptance rates on an H200. That is their benchmark, cited in the throughput post. It is not a denv H200 serving number.

What does transfer is the shape of a deploy. The recipe pins the checkpoint, the engine (vLLM or SGLang), the precision, and the draft setup if you use one. It is tested on H200 before it serves. Weights are cached at the H200 site. A cold site pays the cold start. One measured cold start, on H100, was 12 minutes 46 seconds. Do not assume H200 matches that minute count. It was not measured here.

Failover to the H100 recipe is a second, tested build of the model, not the same bytes on a different card unless the recipe says so. Repair and driver reinstall work the same way.

You pay Nebius, or you pay for the server you joined. denv does not rent the H200. The product is early. Write to anup@denv.ai.