What is a Large Quantitative Model?

For three years, the story of artificial intelligence has been a story about words. Large language models learned to write, summarise, translate and code, and the whole industry reorganised itself around them. But a large share of the world's most valuable decisions are not made in words. They are made in numbers: the binding energy of a drug candidate, the remaining life of a battery cell, the tail risk of a credit portfolio, the stress on a turbine blade at 1,200 °C. That is the territory a new class of systems is built for. They are called Large Quantitative Models, or LQMs.

The short answer

An LQM is an AI model designed to reason over quantities rather than text. In its October 2026 list of top strategic technology trends for 2027, Gartner describes LQMs as modular, hybrid neural-network architectures built to process and analyse quantitative data, and puts them first on the list.

The working definition used across this site is a little broader:

The last clause is the important one. A language model is judged on whether its answer sounds right. A quantitative model is judged on whether its number is right, to a stated tolerance.

Why language models aren't enough

Language models are extraordinary interfaces. Ask one to explain a balance sheet and it will. Ask it to compute the expected loss of a structured product under a thousand correlated scenarios and you are relying on a system whose core mechanism is predicting the next plausible token. Gartner's framing is blunt: using general-purpose language models for numerical work creates accuracy and compliance risk, because the output is a probabilistic guess.

Engineers have worked around this for years by letting the language model call a calculator or a Python sandbox. That helps with arithmetic, but it does not help when the real problem is a model of the physical world — a molecule, a material, an electrochemical cell — that no calculator contains. Someone has to build that model. LQMs are the attempt to build it with the same scaling playbook that worked for language.

Language models are judged on whether an answer sounds right. Quantitative models are judged on whether the number is right.

Anatomy of an LQM

There is no single LQM architecture, and that is part of the point — Gartner explicitly calls them modular and hybrid. In practice, most production systems combine several of these layers:

• A physics or domain prior. Equations from quantum chemistry, fluid dynamics, electrochemistry or financial mathematics, either used to generate training data or embedded as constraints.
• Learned surrogates. Neural networks trained to approximate expensive simulations, so that a calculation that once took hours runs in seconds. PhysicsX, for example, says its platform delivers physics predictions in seconds rather than hours or days.
• Structured knowledge. SandboxAQ has described its LQMs as using neural networks and knowledge graphs rather than the transformer architecture behind most LLMs (VentureBeat).
• An optimisation objective. Instead of "be helpful", the target is a property: a binding affinity, a cycle-life, a drag coefficient, a risk metric.
• A language front end. Increasingly, an LLM sits on top as the interface, translating a scientist's question into a quantitative job and the result back into prose.

Where the data comes from

Language models were trained on the internet. There is no internet of protein–ligand binding energies, and there never will be. That forces LQMs into two strategies, usually combined.

The first is generating data from first principles. If you know the equations that govern a system, you can run them — expensively — millions of times, and train a network to reproduce the results cheaply. In 2025 SandboxAQ released SAIR, an open dataset of roughly 5.2 million computationally cofolded protein–ligand structures, as training material of exactly this kind (AIMultiple).

The second is proprietary measurement data. SandboxAQ's battery work with NOVONIX was trained on more than four million hours of ultra-high-precision cell-cycling data collected over five years. Data like this is scarce, hard to replicate, and is quickly becoming one of the most defensible assets in the field.

What LQMs actually do

Strip away the vocabulary and most LQM deployments do one of three jobs:

1. Screen. Evaluate millions of candidates — catalysts, molecules, alloys, designs — and pass only the promising few to expensive lab work or full simulation.
2. Predict. Forecast a future state from early signals: how long a battery will last, how a part will fail, how a portfolio behaves under stress.
3. Simulate. Run "what if" at a speed that turns simulation from a verification step at the end of a project into a design tool used continuously.

Gartner's phrasing captures the shift: LQMs move decision-making from reactive troubleshooting to proactive simulation. The detailed case studies are in LQMs in the wild.

What they are not

Because the term is new, it is already being stretched. A few boundaries are worth keeping in mind:

• Not a chatbot that is good at maths. An LLM with a calculator is still a language model.
• Not a replacement for the lab. LQMs shrink the search space; physical validation remains the last word.
• Not automatically trustworthy. Quantitative outputs feel precise, which makes bad assumptions more dangerous, not less. Data quality and assumption sensitivity are the field's best-known failure modes.
• Not one company's product. SandboxAQ popularised the term, but the category now spans physics-AI engineering platforms, scientific foundation models and domain-specific forecasting systems.

A working glossary