What the endpoint is for.

These pages are the jobs people show up with. None of them is a different product. Each one is the same cluster, the same recipe, and the same OpenAI-compatible endpoint. The difference is which constraint hurts.

denv does not sell GPUs and does not sell tokens. You pay the provider that owns the machine. The product is early.

We have two published speed results, both one H100, both Qwen3.8-27B, both one request at a time. They are not voice, agent, document, or video numbers. They are linked from the pages that can use them honestly, and omitted where they would be a guess.

• Voice: the GPU has to already be warm.
• Coding agents: long prompts, many turns, a code prompt in the public test.
• Document AI: long context and batch, no published throughput.
• Computer use: a loop of short turns. A failed GPU is the next request, not a silent retry.
• Video: no recipe and no published number. The page says what we will not claim.