Everything we publish is reproducible by design: open weights, open data, open hardware schematics. Browse our programs and the latest results from each lab.
Long-horizon research needs long-lived infrastructure. These platforms anchor every program at the academy.
Trained on 4.1T fully documented tokens. Weights, data mix, and every intermediate checkpoint will be public.
launches Sep 9, 2026 · Apache-2.0A 1.6 m, 34-DoF humanoid with open control stack, built for contact-rich manipulation research in the field.
rev C · open hardware128 superconducting qubits with real-time decoding; remote access granted to 46 academic groups.
gen 3 · open accessIyer, B. Choudhary, R. Choudhary, Beaumont, et al.
Complete training recipe, data audit, and capability account for our flagship open model.
PDF · Code · WeightsSingh, Weiss, Adeyemi, et al.
Logical error rate halves with each increase in code distance — with full calibration data released.
PDF · DataZhou, Roy, Lindqvist, et al.
A dynamics model that stays calibrated through collision and occlusion, enabling long-horizon planning.
PDF · DatasetBeaumont, Krishnan, Osei, et al.
Attribution tooling that links emergent capabilities to specific slices of the training corpus.
PDF · CodeWeiss, Iyer, Tanaka, et al.
A learned decoder that beats matching-based baselines at 1 μs latency on live hardware.
PDF · CodeRoy, Nwosu, Halvorsen, et al.
Design and safety case for a humanoid platform operating alongside people for a full growing season.
PDF · VideoOkafor, Marchetti, Singh, et al.
A continuously refreshed evaluation suite whose items provably postdate any model's training cutoff.
PDF · BenchmarkAdeyemi, B. Choudhary, Bergström, et al.
A chemistry milestone: qubit resources reduced 40× via AI-selected active spaces.
PDF · Data