2.25.1

New Features

• Add center and applyToSubqueries parameters to recency scoring (https://github.com/marqo-ai/marqo/pull/1376)
  • center — A fixed Unix epoch timestamp (seconds) to use as the reference point instead of now(), enabling reproducible recency scores across queries
  • applyToSubqueries — Control which hybrid subqueries ("tensor", "lexical", or both) receive recency boosting in RRF hybrid search

2.25.0

Changes

Triton-based Inference Architecture

Marqo's inference layer has been restructured from a monolithic design into three dedicated components:

• Inference Orchestrator — A FastAPI service that coordinates inference requests. (https://github.com/marqo-ai/marqo/pull/1315)
• Model Management Container — A FastAPI service for managing ML model lifecycles with Triton Inference Server, including model loading/unloading, health checks, and environment variable consistency. (https://github.com/marqo-ai/marqo/pull/1322)
• Marqo API adaptations — The core Marqo API has been updated to work with the new Triton-backed components. (https://github.com/marqo-ai/marqo/pull/1317)

This architecture enables independent scaling and deployment of inference, model management, and search API layers.

Other Changes

• Centralize model registry into a shared components/common package (marqo-common). (https://github.com/marqo-ai/marqo/pull/1356)
• Fix model download auth handling to support public S3 buckets. (https://github.com/marqo-ai/marqo/pull/1352)