Guides · dev · deepseek us hosted

Call Chinese models without sending the prompt to the lab

Published 16 Oct 2026. Updated 16 Oct 2026. Researched 16 Oct 2026.

“US hosted Chinese model” is a host question, not a brand question. DeepSeek the company and a DeepSeek checkpoint running on a US GPU host are different processors. We can show the second only where an endpoint list names the host. We did not pull endpoints for every id. We pulled GLM-5.3 and Kimi K3, because those are the list-price anomalies.

What the feeds named

On GLM-5.3, the lab host is Z.AI at $1.40 / $4.40, the same as the docs. Together and Fireworks were also at $1.40 / $4.40. The cheapest inputs were other companies, with output prices that are not a gift. If your policy says “not the lab”, filter to a named host. If your policy says “cheapest”, you are buying a different company and you should say so in the design doc.

On Kimi K3, Moonshot’s host was $3 / $15. Together and Fireworks matched that rate on the standard tag. Amazon Bedrock publishes its own Kimi K3 card at $3 / $15 on Global CRIS standard. Bedrock is a US cloud bill with its own region choices. We did not open an account to see which region the weights run in. Read the Bedrock region table before you treat the brand as a residency guarantee.

Weights

Kimi K3 weights are public under a licence with revenue triggers. Qwen and DeepSeek have published weights for various generations. A weight file on your disk is the only path where we can say the prompt did not leave the machine, and only if the runtime does not phone home. The local guide stays with smaller Qwen ids for that reason. K3 at the size on the model card is a cluster, not a laptop.

What this is not

It is not a claim that a US host is lawful for your employer. It is not a claim that the host’s subprocessors exclude China. It is a way to stop equating “OpenRouter” with “not China”. The provider tag is the sentence.

Related: safety matrix, OpenRouter versus direct.