LQMs in the wild: five deployments

New technology terms are cheap. Deployments are not. Before the Gartner list made "Large Quantitative Model" a phrase for keynotes, a handful of companies were already shipping systems that fit the description. Five examples show the range — and the common shape — of what LQMs can do today.

1 · Batteries: from months of testing to days

Battery development has a brutal bottleneck: to know how long a new cell chemistry will last, you have to cycle it until it dies, which can take months or years. In 2024, SandboxAQ and battery-technology company NOVONIX trained LQMs on more than four million hours of ultra-high-precision coulometry data gathered over five years.

According to SandboxAQ, the model predicted end-of-life with a mean absolute error of 11 cycles from 40 cycles of data, and around 50 cycles from just two cycles of data — versus around 100 for traditional models. The company estimated the approach could shorten cell development and commercialisation by two to four years.

The lesson is not the specific numbers but the mechanism: an LQM turned a long physical experiment into a short one plus a prediction, and quantified how much it trusted that prediction.

2 · Catalysts, sold like software

Catalysts sit behind more than 90% of commercially produced chemical products, and finding better ones is a classic needle-in-a-haystack search. In June 2026, SandboxAQ announced it would bring its LQMs to the Google Cloud Marketplace, starting with AQCat, a materials and catalyst discovery model that performs adsorption-energy calculations to screen candidates before costly full evaluation.

The target applications read like a decarbonisation roadmap: green hydrogen, sustainable aviation fuel, fertiliser production and plastics recycling. Just as notable is the delivery model. The company says users can reach the models from conversational AI tools without specialised code or infrastructure, and that they can pair with Google's Gemini; it had earlier integrated with Anthropic's Claude. That is the hybrid stack in practice — a language model as the front door, a quantitative model doing the physics.

A language model as the front door, a quantitative model doing the physics.

3 · Drug discovery: binders, affinity and open data

Drug discovery is where quantitative AI has the longest history and the largest prize. Straits Research, cited by SandboxAQ, sizes the global drug discovery market at about $112 billion in 2025, heading toward roughly $187 billion by 2034.

Several LQM-style tools are aimed at its earliest, most expensive stages:

• AQPotency (SandboxAQ) identifies and prioritises protein binders at high throughput and is slated to follow AQCat onto the Google Cloud Marketplace.
• AQAffinity (SandboxAQ) predicts protein–ligand binding affinity from sequence data and is built on OpenFold3.
• SAIR, released in 2025, is an open dataset of about 5.2 million cofolded protein–ligand structures spanning more than a million unique systems — infrastructure for the whole field, not just one vendor.
• Boltz PBC offers a platform that predicts 3D biomolecular structures and binding affinity and designs proteins and biologics.

4 · Industrial engineering at hardware speed

Not every company in the space uses the LQM label, but the physics-AI segment shares its DNA. London-based PhysicsX builds an AI-native engineering platform that, by its own account, runs physics predictions in seconds instead of hours or days. In June 2026 it raised a $300 million Series C at a valuation of about $2.4 billion, led by Temasek, with NVIDIA, Siemens and Applied Materials among existing backers.

Its customers span aerospace and defence, semiconductors, automotive, energy, materials and industrial machinery — every sector where a design cycle is gated by simulation. The company reports doubled recognised revenue and tripled bookings year on year, and is developing what it calls Large Physics Models: larger, pre-trained physics AI models. On the incumbent side, Siemens' Simcenter PhysicsAI brings geometric deep learning into established engineering simulation workflows.

5 · Revenue forecasting

LQMs are not only for labs. Revenue-intelligence company Aviso describes LQMs as the reasoning layer behind its forecasting, pipeline inspection and deal-execution modules. In its design, language models extract signals from emails and meeting notes, and LQMs analyse the resulting numerical data. It is a smaller-stakes example than drug discovery, but a revealing one: the hybrid LLM-plus-LQM pattern is spreading into everyday business software.

The pattern

Put the five cases side by side and a template emerges:

In every case, the model does not replace the expensive step entirely. It decides where to spend it. That is the economic engine of the category — and the reason its value scales with the cost of the experiments it avoids.