Every company makes important decisions from information that was never built for a database. It lives in claims reports, medical notes, source code, system logs, experiment records, images, and conversations.
Most AI models can read this material and generate an answer. But prediction is a different problem.
An insurance adjuster needs to estimate how a claim will develop. A hospital needs to estimate length of stay. An engineer wants to know the latency or accuracy of an experiment before spending the time and compute to run it.
Today, these problems are usually handled one at a time. Teams clean the data, design features, and train a separate model for every outcome. That work is slow, and it breaks down when the useful information is mostly unstructured or when labeled examples are scarce.
We are building one foundation model for this entire class of work.
GQI learns directly from documents, code, logs, images, tables, and other observations. It returns calibrated numeric predictions rather than generated prose. The same architecture supports regression, classification, time-series prediction, and anomaly detection.
Read the model results and research
Sign in to see GQI's current benchmarks, capabilities, and research.
Continue with GoogleWe already have a working model. It can learn a new task from 100 to 1,000 labeled examples and fine-tune in under five minutes. We have validated the architecture on more than 100 tasks across over 40 domains.
Our research precursor, the Quantitative Model, has generalized across scientific and engineering work that normally requires separate specialist systems:
- code, compilers, GPU kernels, and datacenter performance;
- insurance, operations, and other document-heavy decisions;
- scientific experiments, simulation, and health outcomes.
On five diverse prediction benchmarks, GQI-1B outperforms Gemini Pro while costing roughly 170 times less. In other domains, our earlier work has reached state-of-the-art results with substantially fewer evaluations.
This model is the result of three years of research into foundation models for quantitative prediction, including regression language models for code. That work now forms the technical base for GQI.
Our team works across foundation-model research, quantitative modeling, agents, and the systems needed to train and evaluate models at scale.
See the Studio, five-minute demo, or read the documentation.