Self-host DeepSeek V3 on H100.

DeepSeek V3 did not appear anywhere on the site. People still search for how to self-host it. This page is that path, without a made-up benchmark.

DeepSeek V3 is a large open-weight model. A recipe has to say which checkpoint, which precision, which engine, and how many H100s one replica needs. denv will not put a GPU count on this page, because the count depends on the checkpoint and the quant, and we have not published a tested one. Inventing “8×H100” or similar would be a guess.

What you can count on

• The engine will be vLLM or SGLang, pinned by image digest.
• The replica stays in one site. H100 sites in the catalog are Lambda, Nebius, EC2, and servers you join.
• You pay that provider. denv does not sell per-token inference and does not resell the GPUs.
• Health checks, GPU driver reinstall, and replacement of a cloud instance are auto-repair.
• A second site is how you fail over. Both sites need a build that was tested for the GPU it runs on.
• Wide expert parallelism, if the recipe uses it, is the large serving profile: prefill and decode can be split. That is described in concepts and on the llm-d page. It is not turned on by naming the model.

No tokens per second, no cost per million, no customer logo. If you want this recipe, write to anup@denv.ai. The product is early, and the quickstart is still marked launching soon.