Claude Code with an open model on a cloud GPU
Run Claude Code against your own model on a rented GPU
When Claude Code's usage limits stop you, or you want an open model doing the bulk work, run the model yourself on a GPU you rent by the hour.
OpenJet's cloud template serves Qwen3.8-27B over the Anthropic Messages API from an RTX 3090 on Vast.ai or Runpod. Set ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN to the machine's endpoint and token, and Claude Code uses the open model with no proxy and no per-token bill.
Why local wins
- No proxy needed. OpenJet's server answers the Anthropic Messages API directly, so Claude Code connects with two environment variables.
- 121 tok/s on an RTX 3090. Qwen3.8-27B decodes at 121 tokens per second with thinking on, about twice llama.cpp's speed on the same card.
- Hourly, not per token. You pay the GPU provider about $0.20 an hour, with no token metering or usage caps.
Common deployment paths
- Keep working in Claude Code after hitting your plan's limit.
- Run several Claude Code sessions against one rented GPU.
- Use Claude Code on code you don't want sent to a model vendor.
Frequently asked questions
- How do I connect Claude Code?
- Deploy the template, read the API token from the machine's log, then set ANTHROPIC_BASE_URL to the machine's endpoint and ANTHROPIC_AUTH_TOKEN to the token.
- Is Qwen3.8-27B as good as Claude?
- No. It's a 27B open model; frontier models are stronger on hard tasks. It's fast and cheap for routine coding work.
Explore OpenJet
Embed a local model with the Python SDK, run Qwen3.8-27B as a local coding agent, compare local coding-agent alternatives, read the docs, or go back to the homepage.