Deploy DeepSeek R1 on H100.

DeepSeek R1 is the reasoning model people want off a hosted API and onto GPUs they control. denv’s answer is a recipe on H100, not a token reseller.

R1 spends more tokens on the way to an answer than a chat model that answers immediately. That changes cost and latency even when the GPU is the same. This page does not quantify that. We have not published a run. The Qwen H100 numbers are not a stand-in. Qwen3.8-27B in those tests had thinking turned off.

What the recipe has to pin, and what you should ask us to pin:

• the weights revision
• the engine, vLLM or SGLang
• precision and whether a draft model is in the path
• how many H100s one replica uses
• a max output length that matches how you call it, so a long trace cannot fill the GPU by surprise

The replica stays in one site: Lambda, Nebius, EC2, or your server. You pay them. Health checks still matter, because a long generation is the worst time to discover a silent GPU. Failover is for the next request, not for silently resuming a half-written trace on another provider.

The product is early. Write to anup@denv.ai with the checkpoint you mean by R1.