Odysseus setup: Docker and local Qwen2.5 on RTX 3060

This setup guide tests Odysseus, PewDiePie's self-hosted AI workspace, as a ChatGPT alternative. The presenter installs it with Docker Desktop, clones the repository, copies the environment configuration and starts Docker Compose. First login uses the admin account and a temporary password from the container logs; the tour then covers changing that password and enabling 2FA.

The interface includes web search, shell access, plan mode and long-term memory. The presenter first connects OpenRouter and gets a response through MiniMax after another model fails to work. The interface displays API costs and token usage. A research request about ComfyUI produces a visual report that the presenter says draws on 25 sources and uses Tongyi DeepResearch. The report remains accessible in the library.

The cookbook scans hardware and suggests models that fit it. In this test, Docker initially misses the NVIDIA GPU. The presenter pulls an NVIDIA/CUDA image and rebuilds the stack before Odysseus detects the RTX 3060. They download a four-bit Qwen2.5-7B instruct model and set up llama.cpp. Serving through the cookbook fails, so they use the supplied CLI command and add the local model to the interface. The reported speed is about four tokens per second on their 12GB card. This is a practical local LLM walkthrough with setup problems documented, rather than a benchmark of performance across hardware.