Performance estimate

See time to first token, throughput, and memory estimates in seconds — refine as needed.

Model — Hugging Face ID

Tested: Nemotron, DeepSeek V4, Gemma 4, Kimi, ... — type to autocomplete

GPU system

Serving mode:

Based on your configuration — ISL 2048, OSL 128, FP16 KV cache, 1 concurrent users.

Want to change assumptions?

Every number above comes from these. Open a section to tune it — closed sections show their current values.