Memory beyond retrieval

Unlike bidirectionally adapted WAMs, it preserves causal video pretraining, enabling 19.2-second histories to improve control.

Unlike end-to-end audio models, it explicitly plans readable scores before full-song rendering.

A 400M model trained only on synthetic SCM tables beats tuned AutoGluon zero-shot.

Unlike retrieval-style agent memory, it persists latent session state directly, cutting 128K-history storage about 66×.

Unlike most open text-to-image releases, it combines JSON layout control, readable text rendering, and 4K-token prompts in a public 5B DiT with fine-tuning support.

Unlike recent typed-decision repos, it retrofits open LLMs into calibrated classifiers from one prefill, boosting BANKING77 auto-decidable coverage 7.7%→52.0% at ≤5% error.

It pairs 74.0M field-instance masks with a 100-country zero-shot split for cross-country generalization.

Targets dense bilingual document-image editing with region masks and seven text/table edit types.

Unlike recent egocentric manipulation datasets, it adds synchronized 880-taxel full-hand tactile gloves across 100 released human-demonstration hours.