Tag
A survey analyzing 23 open-weight models in the 20B-500B parameter range, covering their architectures and key design choices as of July 2026.
This paper presents a systematic comparison of two geospatial foundation models, TerraMind and THOR, developed under ESA's φ-lab, analyzing how architectural choices like patch size and decoder type affect performance across ten use cases in Earth observation tasks.
A Claude Code skill that autonomously scans codebases for architecture improvement opportunities (e.g., shallow modules), creates delta-aware, analysis-only issues and PRDs, ready for implementation.
NVIDIA utilized late interaction, a form of sparse attention, for an attention-based encoder-decoder to retrieve directly from internal representations.