Running opencode on a self-owned open-model endpoint feels like freedom

A Modal engineer describes the unexpected satisfaction of wiring the opencode CLI to his own inference endpoint rather than upgrading a personal Claude or ChatGPT subscription. The trigger was mundane — he wanted to start a side project but lacked a premium plan on his personal account — so he pointed opencode at Kimi K3, newly available on Modal’s managed endpoints, and had it working in about five minutes.

The piece is less about benchmarks than about the feeling of ownership. Because the endpoint is his and data travels only between his laptop and that endpoint, he describes a sense of control and independence he hadn’t anticipated, comparing the stripped-down experience to opening vim after a heavy IDE. He’s candid that he’s being a bit dramatic, but the reaction is genuine and, to him, surprising given he’d never been an open-source purist.

The takeaway for technical readers: as capable open-weight models like Kimi K3 become easy to self-host on managed infrastructure, the practical and psychological appeal of running your own inference — data locality, no vendor lock-in, low friction — is starting to rival the convenience of hosted proprietary assistants.