Every session starts from zero
Change the conversation and your context, preferences, and working methods need to be explained again.

EVOLUTION IN PUBLIC
From one evolution to continuing growth. Updated with the leaderboard.
Calculated at 196,608 parameters per factory expert evolution. Includes repeated training and illustrative leaderboard records, not actual new parameters.
Updating
LUBAN ZERO · PERSONAL AI
Luban Zero runs on your device. Connect the tools you already use, teach it what you choose, and keep results that pass checks inside your own model.
Current preview · luban0
THE PROBLEM
It can be smart and still behave like an assistant meeting you for the first time.
Change the conversation and your context, preferences, and working methods need to be explained again.
A general answer is not your workflow. The details that matter stay scattered across files and chats.
You use someone else's model, so what you teach rarely becomes durable capability of your own.
THE CORE SOLUTION
Start with one answer, then turn temporary context into model experience that is checked and worth keeping.
Use Q&A, source text, images, or demonstrations. Only explicitly submitted material that passes checks enters the learning flow.
Teacher feedback is a source of material. Independent checks, prior-ability retention, and failure recovery decide what stays.
Use a local API with the AI tools you already use, while keeping your model, materials, and growth records close at hand.
WHY NOW
You do not need to wait for a perfect super-assistant. Start with one real task that matters.
Bring everyday tasks back to your device for more stable access and a clearer data boundary.
Writing, code, references, and decisions form a method. Starting now means you stop rebuilding it from scratch.
Teach one repeatable workflow, verify it once, and make the next use smoother.
WHY KEEP USING IT
Its value is not only how it answers today, but whether it feels more like your assistant next month.
Verified learning results enter your model, so repeated work gradually takes less effort.
See what it has learned, what it still cannot do, and which results were not adopted.
Local runtime, material confirmation, independent checks, and model export keep growth controlled and portable.
GET STARTED
Apple Silicon beta for macOS 26 or later. Extract the app into Applications, open it, and accept the terms. Included models install and start automatically; no further model download is needed.