BWhat changed
Used a lightweight index branch for group-specific key-value block selection and a co-designed block-sparse execution kernel to make long-context attention cheaper.
Scientific scope
Used a lightweight index branch for group-specific key-value block selection and a co-designed block-sparse execution kernel to make long-context attention cheaper. Exact attention applies only inside selected blocks, not equivalence to full dense attention. Paper ablations use a 109B test model; M3 production and hardware speed claims are not interchangeable. M3 is not counted again as a separate attention invention.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Exact attention applies only inside selected blocks, not equivalence to full dense attention. Paper ablations use a 109B test model; M3 production and hardware speed claims are not interchangeable. M3 is not counted again as a separate attention invention.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
- MiniMax Sparse Attention
- Lai et al.: MiniMax group-specific block-sparse attention — official method source
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction