The Era of Vertical AI is Here: Introducing Luna Crescent – Luminance’s Proprietary Legal Intelligence Model  

The Era of Vertical AI is Here: Introducing Luna Crescent – Luminance’s Proprietary Legal Intelligence Model  

By Dr Graham Sills,Co-Founder and Director of AI

The first wave of enterprise AI was defined by access. As foundation models became increasingly powerful and widely available, businesses rushed to integrate them into products, workflows and entire business processes. Suddenly, every software company had access to the same remarkable technology. 

But as AI matures, a new reality is emerging. General-purpose models are getting better and better. Yet some problems are simply too specialized to be solved optimally by models designed to solve everything. 

Legal work is an example of this. Contracts are not just another form of text. They are filled with nuance, complex drafting conventions and subtle distinctions that can materially change meaning, risk and commercial outcome. Understanding them requires more than language understanding. It requires legal understanding. 

It’s why the next phase of enterprise AI will be defined by vertical models: specialist AI built for specialist work. At Luminance, we’ve been building towards that future for some time. And today, we’re announcing that future in the form of Luna Crescent – Luminance’s proprietary legal intelligence model. 

Why did we build Luna? 

Luminance has been building Legal-Grade™ AI for over a decade. In that time, our platform has been exposed to over 220 million verified legal documents. This powerful corpus offers something general-purpose model providers cannot replicate: a deep, ground-level understanding of how contracts are actually written, negotiated and managed in practice. Legal language is not a subset of general language. It operates according to its own logic, conventions and risks. 

As foundation models became increasingly capable, the opportunity became clear: combine the power of general AI advances with a specialist intelligence layer that only years of legal data and domain expertise could produce. That’s the principle behind Luna.  

Luna is not a replacement for foundation models. Luminance’s multi-model architecture continuously evaluates and combines the best combination of models for any given task. Where foundation models deliver the strongest results, we use them. Where a specialist legal model performs better, we use Luna. The goal is to combine the best of both for the optimal outcome. 

Where does the power come from? 

Luna was designed with a clear goal: to create a specialist legal model optimized primarily for accuracy and legal understanding.  

Its training draws on Luminance’s extensive, high-quality dataset – specifically a carefully curated selection of challenging examples spanning around 3.4 million legal concepts and data points. Importantly, the breadth of the training data reflects the complexity of legal work: contracts from virtually every industry, including highly complex agreements, difficult file formats, obscure drafting styles and diverse legal language. This enables Luna to perform effectively across the full range of contracts encountered in real-world legal scenarios.  

Reliability was a core design principle throughout. Rather than generating plausible-sounding answers, Luna has been specifically trained to prioritize faithful extraction from source documents. When information is not present, it is designed to identify that absence rather than invent an answer. We know instilling this level of trust is what matters most for legal professionals. 

What does this mean in practice? 

Ultimately, the value of any AI comes down to results. Luna is already operating within the Luminance platform, helping power the contract intelligence capabilities used by customers around the world. 

To validate performance, Luminance’s Cambridge-based R&D team developed ContractIQ Bench, a proprietary benchmark consisting of 189,000 manually annotated and reviewed data points which assess a model’s ability to interpret key provisions within legal documents, such as liability caps, termination for convenience, and confidentiality obligations. Using ContractIQ Bench, Luminance tested Luna against leading general-purpose models on held-out legal documents and concepts not included in Luna’s training data, alongside blind evaluations by legal experts.  

Luna achieved 5% higher accuracy on contract understanding tasks and was consistently preferred by legal experts for its ability to identify contractual concepts and extract key information from complex clauses.   

The advantage is clearest in work that general-purpose AI was not purpose built to solve: identifying non-standard clause constructions, detecting concepts expressed in unfamiliar language, classifying obligations buried in dense boilerplate and flagging what is absent as well as what is present. Luna understands legal language as a system of obligations, rights, risks and commercial consequences, rather than as another category of general text.  

Luna offers this specialised legal intelligence while also processing information quickly. The model generates text nearly 4 times faster than generalist tools, reaching 200-400 tokens per second. Work that takes 30 seconds on a general model takes under eight on Luna. Multiplied across an enterprise contract repository, those seconds become hours of recovered time and faster business decision-making.  

The benefits extend beyond quantitative metrics. Model ownership also delivers meaningful security advantages. Luna is deployed within Luminance’s own AWS environment, reducing reliance on external model providers and limiting data exposure to additional sub-processors. This gives Luminance greater control over performance, reliability and deployment, while helping protect customers from disruptions related to third-party availability, pricing or access.  

The next phase of enterprise AI  

So what’s next for enterprise AI? There’s no doubt that foundation models will continue to improve, and remain an important part of any technology stack. But the greatest value will increasingly come from models built for specific domains, trained on specialist data and optimized for specialist work. 

For legal AI, that’s exactly what Luna represents – a specialist intelligence layer built for the future of contract intelligence. That’s the standard Luminance has spent a decade building toward.