Lumbridge

Train in the cloud. Run on hardware you own.

Open-source infrastructure for running your own models on your own hardware.

Six projects. Four published under Apache‑2.0. Nothing here requires an account with us.

The projects

• ComputeGoverned compute for unified-memory hardware. Admission control that refuses to overcommit, and a 1 Hz memory watchdog.Apache-2.0
• BenchQuality and serving performance measured against one pinned target: model × host × quant × serving config.Apache-2.0
• ArenaRL environments on the verifiers v1 spec. Each is graded on a held-out slice the model never read.Apache-2.0
• SimulateA three.js engine over pure-data city packs. Sun, moon and city lights computed from the clock.Apache-2.0
• ForgePost-training as a content-addressed build: base checkpoint, data mixture, adapter and merged shards under one signed manifest.Not published
• StreamA headless broadcast rig. One encode of a live display fans out to every destination at once. Runs in production.Not published