What are LQMs and why they matter

While Large Language Models (LLMs) like GPT-5.5 or Claude generate text based on statistical patterns from billions of web pages, they have no real understanding of the physical world. If you ask them which molecule will bind to a specific catalyst, they'll answer based on texts they've read — not based on the laws of chemistry.

Large Quantitative Models (LQMs) from SandboxAQ solve this problem from the ground up. Instead of training on texts, they're trained on data from quantum-mechanical simulations. Specifically, they use Density Functional Theory (DFT), molecular dynamics, and other computational methods that model the behavior of atoms and molecules according to physical laws. The result is a model that doesn't hallucinate — because its answers are backed by hard physics.

"LLMs generate content. LQMs generate results," summarizes the company's approach, as stated by CEO Jack Hidary, who introduced the concept on May 18, 2026, during a live broadcast on CNBC's Squawk on the Street.

Three-layer architecture: data, models, agentic workflow

SandboxAQ builds its LQMs on a proprietary three-layer infrastructure:

1. Data that the company generates itself. While most AI models rely on public datasets, SandboxAQ generates its own training data using physical simulations. This allows it to work even in areas where experimental data is missing — for example, with entirely new chemical compounds. "We're not limited by existing data. When public databases fall short — and with new chemistries they often do — we don't stop," the company explains in its integration announcement.

2. Proprietary models with open collaboration. The portfolio includes AQCat for catalyst screening (20,000× faster than traditional DFT calculations), AQVolt for discovering materials for solid-state batteries, and AQAffinity for predicting protein binding affinity (1,000× faster than FEP simulations). The company also publishes openly — the AQAffinity model is built on the open-source OpenFold3 and continues to improve on it.

3. Agentic workflows that make decisions. The third layer orchestrates complete "design–test–make–decide" loops. Agents plan, iterate, and present results, while the scientist remains in control at key checkpoints.

Claude as a gateway to the world of quantitative AI

The key announcement from May 18, 2026, is the integration of LQMs with Anthropic Claude via the Model Context Protocol (MCP). MCP is an open protocol that Anthropic developed for connecting AI assistants with external tools and data. SandboxAQ is one of the first to use it to connect specialized scientific models to a conversational interface.

In practice, this means a scientist can write to Claude in natural language: "Calculate the adsorption energy of hydrogen on a platinum surface" — and Claude will send the query to AQCat Adsorption Spin, which returns a physically accurate result. Claude then interprets the result in an understandable way.

The first integrated model — AQCat Adsorption Spin — is already available through a waitlist. It's a spin-aware model for heterogeneous catalysis that enables rapid identification of the most promising catalyst candidates before expensive laboratory testing.

AQCat25: a catalysis breakthrough published in Nature

In early May 2026, the SandboxAQ team published the AQCat25 model in the prestigious journal npj Computational Materials (part of the Nature portfolio). It's a dataset of 13.5 million DFT calculations across roughly 47,000 adsorbate–surface systems, generated with significantly higher accuracy than previous large-scale datasets for heterogeneous catalysis.

What makes AQCat25 exceptional:

• Spin polarization for 12 elements including iron, cobalt, and nickel — the model "sees" magnetic properties that govern the behavior of many industrially important catalysts.
• Expansion of chemical space with elements like lithium, barium, lanthanum, cerium, magnesium, and fluorine — closer to real-world catalyst formulations.
• Better success rate at finding the global minimum — the model more often finds the physically correct adsorption energy minimum, meaning fewer false-positive candidates for expensive lab tests.

Both the dataset and models are publicly available. For R&D teams, this means the ability to start analysis immediately.

Impact on pharma and materials research

SandboxAQ isn't just targeting academia. The company's customers include Sanofi, Dow, Procter & Gamble, the U.S. Air Force, and the U.S. Department of Health and Human Services. In drug discovery, the company is tackling a decade-old problem in computational chemistry — predicting the free energy of ligand binding to a protein without needing to know the crystal structure. The AQAffinity model handles this roughly a thousand times faster than the current gold standard.

For the battery industry, the AQVolt26 model (published in April 2026) accelerates the search for solid-state battery materials — a key technology for electric mobility, where Europe is heavily betting on its own manufacturing capacity.

What this means for Europe and Czechia

For the European pharmaceutical and chemical industry — including companies like Sanofi (with branches in the EU), BASF, or Bayer — LQMs represent a potentially transformative tool for accelerating R&D. Shortening the new molecule discovery cycle from years to months would mean billions in savings.

For the Czech environment, the openness of the ecosystem is key. The MCP protocol is open-source, and LQM datasets like AQCat25 are publicly available. Czech academic institutions — from the University of Chemistry and Technology (VŠCHT) through the J. Heyrovský Institute of Physical Chemistry to CEITEC — can thus test and adapt the models without licensing fees. Claude and LQMs don't yet support Czech as an input language for scientific queries, but prompting in English is not a barrier in these fields.

The investment dimension is also noteworthy: SandboxAQ's backers include Yann LeCun, Turing Award winner for deep learning, and the venture fund of former Google CEO Eric Schmidt. The company, which was founded in 2022 as a spin-off from Alphabet, is earning the trust of the field's top minds.

The future of conversational science

The integration of LQMs with language models points to the direction scientific AI is heading: specialized models for specific physical domains, accessible through universal conversational interfaces. SandboxAQ promises that more models will follow in weeks, not months.

For researchers, this means the end of an era where they had to be experts in both DFT calculations and Python programming. Instead, they can simply ask a question in natural language and let AI handle the computation. The scientist can then focus on what they do best: asking the right questions.