Measuring Cold Starts on a 27B Hugging Face Endpoint
Oct 1, 2026·8 min read
I deployed Qwen3.8-27B on a dedicated Hugging Face Inference Endpoint, scaled it to zero three times, and measured what cold starts actually cost: 6.6–10 minute boots, 36–50 hard 503s, and roughly $0.28–$0.42 of billed startup per wake.
