Model intelligence
Exploring architectures and training methods that strengthen reasoning, learning, and generalization. We study how models can solve unfamiliar problems and build on what they know.
Research
Advancing intelligence. Expanding human potential.
We pursue breakthroughs in AI to expand what humanity can understand, create, and achieve. Our ambition is a future where intelligence is no longer a barrier to progress, and its benefits are shared by all.
Research focus
Models and agents. Learning together.
We connect models, agents, and efficient compute to build AI that reasons more deeply and acts more reliably.
Agent autonomy
Developing systems that turn understanding into action. Our research focuses on planning, memory, and self-correction so agents can pursue complex goals over longer horizons.
Compute efficiency
Making greater capability possible with fewer resources. We investigate efficient architectures, low-precision computation, and faster inference to reduce the cost of training and using advanced AI.
Long-term research
Exploring the foundations
Our long-term research asks how intelligence can help humanity reach further.
New architectures
We explore new architectures, richer memory, and training methods that could help AI tackle harder problems and expand what people discover and create.
Efficient intelligence
We investigate more efficient pretraining and inference so advanced intelligence is accessible to more people.
Corepass Models
Our model program
Our model program connects research in architecture, training, and inference with what we learn from real agent workloads. Each model explores a different balance of capability, speed, and compute efficiency.
- Scout
- In production
- Small, fast models for focused tasks such as code search and context retrieval.
- Voyager
- In development
- Our main coding and reasoning model, with a focus on reinforcement learning for agent tasks.
- Horizon
- Planned
- A larger model for the most difficult reasoning and agent tasks.
Research infrastructure
From experience to intelligence.
Vera
Automated RL data generation
Vera turns real tasks into learning opportunities: reproducible environments, agent rollouts, and calibrated verifiers. It automates the generation of training and evaluation data for reinforcement learning, grounding progress in work that can be checked.
Forge
Reinforcement learning at scale
Forge is our reinforcement learning framework for connecting tasks, model inference, verified rewards, and training. We are developing it to make experimentation more efficient and turn experience into more capable, reliable models.
We bring model researchers, systems engineers, and product builders together to turn research into capabilities people can use.
Join our team