Patent Strategist

Patent Strategist is an open AI model tuned to reason about patents. A patent is the legal cover that protects an invention, and the work around it, like writing claims and answering the patent office, is slow, careful, and full of private client text. This model does that reasoning fully offline, so none of that text ever leaves your machine.

What it can do

It helps with the day to day of patent work: shaping the claims that define an invention, drafting replies to the patent office that point to the right rule, weighing how close earlier inventions are, and thinking through licensing deals. It is built on DeepSeek R1, a strong reasoning model, and made smaller and sharper with NVIDIA’s NeMo toolkit so it fits on a single small desktop.

How well it works

We scored it on our own 200-question test set, in three setups: with no help, with the ability to look things up, and when handed the exact rule it needs. On multiple-choice patent law it climbs from 63 to 95 percent as it gets more help, and it nails the structure of an office-action reply every time. The numbers above tell the full story.

It is not perfect. A couple of made-up rule citations slipped in from the training data, so treat it as a sharp first draft and always check its sources.

How to run it

Download the GGUF files (the ready-to-run format) and run them with llama.cpp on a Spark-class desktop, a small AI machine with 128 GB of memory. The Q5_K_M build is the sweet spot: about 35 tokens a second, with quality close to the largest build. Builders who want to fold the patent skill into their own model can grab the LoRA adapter, a small training patch, instead.