Public KG record
Heavily Compressed Attention
Recorded overview
Caches 1 entry per 128 tokens, used in 37 of 43 layers with 512-dim heads.
An evidence-backed entity from the project knowledge graph; detailed tool metadata is still being added.
Evidence coverage
The KG currently records 6 known information dimensions and has 13 undocumented dimensions, linked to 2 public source articles. These counts describe record coverage, not model performance or verified fit.
Related public records
- Hybrid Attention — integrates
- ERNIE 5.5 Flash Base — uses