Oryx
Summary
Oryx is the sequence-axis hybrid model introduced by Multi-Mixer Models: Flexible Sequence Modeling with Shared Representations. It lets a block switch between softmax attention and a linear recurrent mixer across different sequence chunks while sharing more than 90% of parameters across mixer modes.
Role In The Wiki
Use this page as the object card for the Oryx method. The source page carries evidence, limitations, and TSFM agenda mapping.
For the wiki’s long-context and fixed-FLOPs discussions, Oryx is upstream architecture evidence for budgeted mixer selection. It suggests a time-series model might combine exact attention over selected spans with compact recurrent updates over ordinary spans, but it does not provide direct evidence for numeric time-series state, event streams, or action-conditioned world models.
Evidence
Official Artifacts
- Preprint: arXiv 2605.28769
- X provenance snapshot in local raw layer:
papers/oryx-2026/x_thread_kevinyli_atrost3122_2067658710810603549.md - Official code or model checkpoints: none found at ingest time.