Kimi K3 rebuilt Google Maps 3D in 90 minutes, two prompts

Someone handed Kimi K3 two prompts and got back an interactive 3D globe: real building extrusions with height data, shadows that track the sun, free camera pitch and zoom, clickable landmark cards, and smooth flyTo transitions between any two points on Earth. Full Google Maps 3D behavior, clean modular code, env-based tokens — in about 90 minutes. It's live at 3dmap.kimi.page.

The demo lands because of what's under it. Kimi K3 is Moonshot AI's new 2.8-trillion-parameter model — the largest open-weight release to date — built for long-horizon coding and agent workflows with a 1M-token context window. On Vals AI's overall benchmark it ranked second, trailing only Fable 5 and edging out GPT-5.6 Sol, which is exactly the tier the post is bragging about undercutting on cost.

The real hook is the license: full weights drop July 27 under a Modified MIT license. A near-frontier coding model you can run yourself changes the math on what a one-off 3D map build should cost.

Why it matters: frontier-grade agentic coding is about to be open-weight and cheap to self-host.