punyaslokdutta wants Mike running on your own hardware

A proposal to let Mike answer questions using open-weight models running locally, with no hosted AI account required.

infrastructurechat-ui

Most Mike forks lean on a hosted AI provider, which means an account, an API key, and your prompts leaving the building. punyaslokdutta sketched an alternative: point Mike at a model running on your own machine or servers instead. The work bundles a few distinct pieces:

  • Local model support - Mike can talk to open-source model runners (Ollama and vLLM, tools for hosting AI models yourself) with no hosted account or key needed.
  • New model choices - Qwen 3 8B and 14B, two open-weight models, show up as selectable options in Mike's picker.
  • Answers that stream - text now appears word by word as the local model generates, instead of Mike sitting blank while a slower local model thinks.
  • A documented setup - instructions for standing the whole thing up locally to test end to end.

Worth noting: this was submitted as a proposal and closed without being merged, so it is a blueprint rather than a shipped feature.

So what Anyone weighing Mike for data that can't leave their own infrastructure should read this as the starting point for a fully self-hosted deployment.

View this fork on GitHub →

Spotted something wrong? Or know the PR text has fresher detail than the writeup above?