Deploy Llama 3.1 8B on H200.

H200 is a lot of GPU for Llama 3.1 8B. The reason to do it anyway is the machine you already have, usually a Nebius H200, or a server you joined. denv will not rent you the H200.

The deploy is the same shape as the H100 8B page: a recipe pinned to this GPU type, health checks, an OpenAI-compatible endpoint. The recipe has to have been tested on H200. An H100 recipe is not a shortcut.

There is no published tok/s. DFlash’s own draft-acceptance numbers were measured on an H200 in the drafter’s benchmarks, and they are not a denv Llama result. They are cited only in the 213 tok/s post, for Qwen.

MIG and DRA are the practical question on this page. If 8B is one of several models, give it a slice. If it is the only model and you want the whole card, say so in the recipe and accept that you are paying for idle silicon.

Failover to an H100 site is allowed only for a build of the same model that was tested on H100. Otherwise you are failing over to a different system and calling it the same endpoint.

Write to anup@denv.ai. The product is early.