Private testing
Open-weight LLM inference · Japan
Open-weight language models on low-power hardware, behind an OpenAI-compatible API. We write our own kernels and tune them for tokens per watt. Nothing you send is stored, logged or used for training.
01 · Offer
One API, open models, nothing kept.
Chat completions with streaming and token usage. Point an existing OpenAI client at our base URL and it works.
We serve models whose weights are openly published. The /models endpoint lists each one with its price.
Prompts and outputs are not stored after the request, not logged, and never used for training.
02 · Efficiency and privacy
Tokens per watt.
We write our own kernels and measure every change against the hardware's limit. Less power per token means less heat and a lower running cost.
Content exists only while your request runs. We keep the meter, not the message.
Requests are processed on our own servers in Japan.
Models, prices and uptime will be listed here when the public endpoint opens.
03 · Contact
Privacy and data requests: privacy@coldservelabs.com
Cold-Serve Labs