The harness is the product.
It runs on GLM-5.3 and MiniMax-M3. The prompts, the tools and the research pipeline are all hand-built — and they are what let it hold its own against harnesses running far larger models.
The whole of it
- Models
- GLM-5.3 and MiniMax-M3, switched at runtime
- Research
- splits the question, reads the primary sources, then argues with its own answer
- Reads
- PDFs, spreadsheets, and the web — with the sources kept
- Makes
- answers, pictures, songs
- Hands off to
- six specialists — songwriting, image prompts, channel growth, frontend design
- Remembers
- what it learns, unless you switch it off
- Keeps going
- long chats compress instead of hitting a wall
- Runs on
- one machine, one GPU, in a house
- People
- four
Lately
- Aug 30the interface steps back from the conversation
- Aug 30spreadsheets go in, and a big one arrives as a digest
- Aug 30a switch for chats you would rather it did not learn from
- Aug 30poke the pet, and it likes you
- Aug 30leave the pet alone and it falls asleep
- Aug 30the mic makes a sound opening, and the same sound backwards closing
- Aug 30live dictation — text streams in and revises as you speak
- Aug 27a survey of the last month was 80% songs with nothing saying so
- Aug 26deleting one memory tore down the whole tab
- Aug 15a song can have versions of itself
- Aug 13make the minute a song takes visible, and survivable
- Jul 22stop the model claiming a song it never generated
Every line above is a real commit message. 951 tests, 23 tools the model can call, and most of it written late at night.