Stilta Reinvents Patent Work With NVIDIA
The technology
Model: NVIDIA Nemotron 3.5 Lightning, post-trained with supervised fine-tuning on real litigation claim charts and reinforcement learning aligned to patent examiner decisions. Deployed on and powered by Microsoft Azure.
Results: Outperforms every open and closed model tested on European Patent Office examiner decisions, lifting the label score from 34.5% to 57.6%. On one NVIDIA H100 GPU, it runs 5,000 claim charts per hour, 8–13x faster than open reasoning models, at $0.65 per 1,000 comparisons versus $132 for the leading frontier model.
Impact: ~200x more cost-efficient at scale than frontier models, at higher accuracy. A full 10,000-family portfolio can be analyzed every month for about $1,600 a year, instead of $320,000.
Protecting innovation has never been more important. Patents are the legal, protective layer of the world’s most valuable inventions, and the race to secure and leverage them is accelerating: global patent filings hit a record 3.7 million recently, and in AI, semiconductors, and energy, companies and nations are staking out stronger positions. The billion-dollar disputes that follow can decide the fate of companies and entire industries. Yet the work of understanding and defending those positions has stayed slow, expensive, and limited by human review.
From One Patent to the Whole Portfolio
Stilta set out to change that. The company began with the hardest atomic unit of the problem: analyzing a single patent to identify and defend against infringement. It mastered that task and built a platform around it, today used by leading enterprises such as Maersk, and law firms such as Mannheimer Swartling.
Now Stilta is taking that precision from one patent to every patent. Enterprises do not hold one invention; they hold thousands, and until now they have been judged by count, not content. Yet a portfolio’s real value, and its real exposure and protection, lives in what the patents actually claim and how they connect to each other, to the products on the market and to the prior art scattered across the world. Mapping them is not a bigger version of the old task. It means reasoning across millions of connections at once, a problem of a magnitude that has been out of reach, and one that has left an entire industry operating with a systemic blind spot.
“Patent portfolios have always been measured by their size. Built on NVIDIA, we can now analyze what a portfolio actually protects, and expose the risks and opportunities before a competitor does.”
Oskar Block, CEO and co-founder, Stilta
A Task Built for a Scalable and Specialized Model
At the core of every Stilta analysis is one repeatable task: checking each element of a patent claim against a product or earlier publication and citing the passage that proves it. It is narrow, well-defined, and run millions of times across a portfolio, which makes it an ideal target for a specialized model rather than a general-purpose one.
Stilta trains the Nemotron 3.5 Lightning model on NVIDIA accelerated computing and benchmarks it against leading open and closed models on a public dataset of European Patent Office examiner decisions before anything reaches a customer. After post-training, the model lifts the label score from 34.5% to 57.6% and the rationale rubric from 47.7% to 59.3%, above every open and closed model tested, including Claude Opus 5 and GPT-6 Luna, while running 8–13x faster than open reasoning models such as gpt-oss 120B and Qwen3.6 35B-A3B. It will be deployed on Microsoft Azure, giving customers the enterprise-grade scale and security that confidential IP work demands.
Why it Matters in IP
Twenty years ago, cybersecurity was handled by specialists after a breach. Today it is a board-level discipline, running continuously on software that watches every system. IP is on the same path. As innovation becomes the most critical asset companies compete on, protecting it can no longer rest on a few experts reviewing one patent when a dispute arrives. It needs the same data-driven discipline, only here what is being protected is not data, but innovation.
That is what portfolio-scale analysis brings: the picture of where a company is strong, where competitors are closing in and where it is exposed moves out of people’s heads and into software grounded in the actual data, surfacing the risks and opportunities that were infeasible to find before.
The work is ongoing, but early access is now open to a limited number of leading enterprises and firms. Learn more by contacting the team.
About Stilta
Stilta is an AI platform for patent analysis work, redefining the analysis at the core of patent infringement, invalidity and freedom-to-operate. Used by Fortune 500 enterprises and leading IP and law firms.
