Reasoning with executable traces

Adds product-quantized next-concept prediction to autoregressive decoding, matching OLMo-3-7B loss with 51.3% tokens.

Unlike larger post-trained specialists, it trains an open 7B from scratch to 256K context, ranking first across 14 reasoning benchmarks and reaching 4.2x faster 16K pretraining time-to-loss.

Reached IMO 2026 gold purely with natural-language generate-verify-refine search, without formal tools.

Unlike GPU-focused simulators, it exposes 4096-way CPU MuJoCo batching with direct NumPy state access.

Unlike tool-use agents needing finetuning, it improves OSWorld 2.0 partial credit 71.97→78.98 using frozen-memory exploration across 82 tasks.

Unlike agent runtimes, it ships a 744B MoE model with leading tool-use benchmarks.

Unlike smaller or single-turn tool-use sets, it releases 484k executable multi-turn trajectories with tool schemas, observations, errors, retries, and final answers.

Unlike text-to-motion datasets, it canonicalizes 7,430 heterogeneous humanoid, animal, and object skeletons with topology and orientation annotations.

Provides 250k aligned reasoning traces across 44 non-English languages, unlike mostly English-only reasoning datasets.