One model, one GPU type.

A recipe is a tested bundle for one model on one GPU type: engine, version, precision, arguments, serving profile, autoscaling, alerts, and the repair policy. It is pinned by image digest and weights revision. The homepage calls this the catalog. These pages are the catalog you can link to. They are not a promise that every pair is live today. denv is early. If a recipe is not in the catalog yet, write to anup@denv.ai or bring the weights. A custom recipe is tested on your GPU type before it serves.

Llama 3.1

• Llama 3.1 8B on H100
• Llama 3.1 70B on H100
• Llama 3.1 8B on H200
• Llama 3.1 70B on H200

The only Llama string on the homepage today is the sample llama-3.1-8b-instruct in the Python snippet. That is an example model name, not a benchmark.

Qwen

• Qwen3 on H100
• Qwen3 on H200

The measured model is Qwen3.8-27B, not every Qwen checkpoint. Read the benchmark before you quote a number.

DeepSeek

• DeepSeek V3 on H100
• DeepSeek V3 on H200
• DeepSeek R1 on H100
• DeepSeek R1 on H200

No denv tokens-per-second figure exists here for DeepSeek. The pages say how a deploy works and stop there.

Engines are vLLM or SGLang. GPUs you do not see above: B200, A100. You pay the provider directly.