The Research division pushes the frontier of efficient computing and logic models, building systems that reason more while using less, then publishing the methods, evaluations, and weights.
Architectures and training methods that get more capability per FLOP, so strong models run on far less hardware.
Models that reason in structured, verifiable steps, not just fluent guesses, so their conclusions can be audited.
We release weights, code, and the evaluations behind them, so results are reproducible rather than asserted.
Efficient computing paired with logic and reasoning, optimized as one system rather than for isolated benchmarks.
We document data sources, curation methods, and training procedures. If you build on our models, you should know where they came from.
The same research powers our products, deployments, and advisory work, so what we learn in one place compounds everywhere.
We're hiring researchers and engineers, and we collaborate with teams pushing the same frontier.