DocsHelp
FAQ
Common questions about Osarai: what you pay for, privacy, which models and tools work, and how instances behave.
What am I paying for?
GPU time. A running instance costs its hourly price, whatever it serves; tokens are free. See Billing.
Does anyone else use my GPU?
No. Each instance is a machine rented for you alone, and only your API keys reach it. Requests do not go to a shared pool.
Do you store my prompts?
No. Prompts and replies pass through to your instance and back without being stored; Osarai records only the time, model, endpoint, key, status, token counts and duration of each request, for your usage charts and billing. See the Privacy Policy.
Can I use Claude models?
No. Osarai serves open models. Claude Code works because Osarai speaks the Anthropic Messages format, but the model behind it is the one on your instance, such as qwen3.8-27b.
Which model should I start with?
qwen3.8-27b on an RTX 5090: reliable tool calls, a 128K context and strong code skills. For Codex CLI, use gpt-oss-20b. See Models and limits.
Which tools work?
Anything that speaks the OpenAI Chat Completions, OpenAI Responses or Anthropic Messages API. Guides cover opencode, Claude Code, Codex CLI, Crush, Qwen Code, Aider, Goose, Kilo Code and Continue, the Cline, Roo Code, Continue and Zed editors, and the OpenAI, Anthropic, Vercel AI, LangChain and LiteLLM SDKs.
Why does a deploy take a while?
Each deploy rents a fresh machine, which downloads the model (tens of gigabytes) before it can load it. The instance page shows each step while you wait.
What happens when I am not using it?
An instance stops by itself after a stretch without requests (60 minutes by default), and billing stops with it. Deploy again from the dashboard when you need it.
Can I run more than one model?
Yes. Deploy several instances; each request goes to the instance whose model matches model, or the one you pin with X-Instance-Id.
Is there a rate limit on tokens?
No. Each key has a requests-per-minute limit (120 by default, up to 600); how many tokens you generate depends only on your GPU.
Can I use my own model?
Admins can register other models; ask through Support if you need a model that is not in the list.