Model learning
Our model research explores architecture, pretraining methods, post-training, and reinforcement learning. We study how these approaches can develop stronger reasoning, coding, and tool-use capabilities.
About Corepass
Research,put to work.
Corepass is an AI research and engineering lab based out of Menlo Park, California. We develop models and agent systems together, turning advances in reasoning, learning, and compute efficiency into AI that can carry out real work.
Our purpose
Intelligence in
service of humanity.
We envision a future where intelligence is no longer a constraint on human potential. Where more people can discover, create, and solve problems once beyond our reach.
We advance AI in service of humanity expanding human agency, supporting human flourishing, and helping the benefits of progress reach people everywhere.
Research & engineering
Better models.
More capable agents.
We study the model and the system around it as one problem. Better reasoning matters. So do the context a model receives, the tools it uses, the feedback it learns from, and the compute it takes to finish.
Explore our researchAgent systems
We investigate context retrieval, planning, memory, and verification for long-running tasks. Remy puts those ideas to work and provides a setting for evaluating what succeeds, what fails, and what to improve next.
Compute efficiency
Our systems research spans low-precision computation, multi-token prediction, memory, and kernel optimization. We measure progress in the quality, time, and total cost of completing a task.
Our approach
A short path from hypothesis to evidence.
Research grounded in real tasks
Let difficult coding and knowledge-work problems guide our questions. Use verifiable outcomes to distinguish an apparent improvement from a useful one.
Experiment, measure, iterate
Move quickly from an idea to a prototype, evaluate it, and learn from failures. Bring promising results back into the product and check whether the gains hold.
Optimize the whole system
Study model quality, context, tool use, latency, and compute together. Account for failed attempts and retries when measuring the cost of useful work.
Our team
Research and product, in the same room.
Our team brings together researchers, training and inference engineers, evaluation specialists, agent engineers, and product builders. The people investigating model behavior work closely with the people shipping the systems those models run in.
We value clear hypotheses, careful measurement, and the willingness to change course when an experiment disagrees with an idea. We want research to improve products, and product experience to raise better research questions.