Zember Furnace — local AI for 24 GB graphics cards
Granite 4.1 30B on your own PC, for 24 GB graphics cards. Reads every page of a contract, cites the page, uploads nothing.
Check your computer before you buy
- A graphics card with 24 GB of memory or more
Known to fit: RTX 3090, RTX 4090, Tesla P40 24GB, Tesla M40 24GB. - Windows
Task Manager → Performance → GPU → Dedicated GPU memory shows what your card has. - Each model is a separate product
The key opens Zember Furnace and only Zember Furnace. Pick the one for your card.
What it does
- Reads the whole contract
For a summary, the deadlines, the parties, the amounts or the risky clauses it reads every page, part by part. - Cites the page for every point
Like (p. 7), so you can check it before you rely on it. - Catches the clause that changes another
The app finds every clause that replaces or amends an earlier one and quotes it exactly, so the value that applies now comes first. - Nothing leaves your computer
It only listens on this machine. After the brain is installed you can unplug the internet.
What we measured
- Contract test: 12 of 12 answers right
A 3-page contract built to trip it up (the real late-payment penalty in the last clause, wrong figures on the way, a later clause replacing an earlier one). Four questions, each asked three times, through the very app in the box. - 5.1 tokens a second on our NVIDIA GeForce RTX 5070 (12 GB) + 64 GB RAM
That card is smaller than the 24 GB this model is for, so part of it sat in RAM; on the right card it goes faster.
What is inside
- Granite 4.1 30B Q5_K_M
19.1 GB, Apache 2.0. It comes straight from the people who made it, at a fixed version, on the first start. - The app, the engine and OCR for scanned pages
About 135 MB to download. The manual inside also shows how to install the brain without internet.
It is a reading assistant, not a lawyer: check every point against the page it cites.