Summary
MLflow 3.9.0 introduces an in-product MLflow Assistant chatbot and a Trace Overview Dashboard for GenAI experiments, enhancing debugging and performance insights. The AI Gateway is revamped for direct tracking server integration, alongside new LLM judge features for online monitoring and custom prompt building.
Summary generated by brickster.ai. For the full changelog and any code/binary attachments, follow the GitHub link above.
More from mlflow/mlflow
MLflow 3.16.0
MLflow 3.16.0 makes the redesigned trace explorer the default interface, introducing natural-language custom trace views via the MLflow Assistant, session grouping for multi-turn conversations, and span links. The update also adds Unity Catalog model service support for built-in evaluation judges, per-user AI Gateway budget policies, and fail-closed authorization by default.
MLflow 3.15.2 introduces immutable evaluation dataset versions and a scorer_ensemble primitive for combining scorer results. The patch also fixes a Databricks telemetry deadlock and addresses issues in MemAlign aligned judges and runs.status constraint handling.
MLflow 3.15.1
MLflow 3.15.1 fixes version parsing issues on Databricks Serverless and corrects env_pack behavior on ARM client images. Scorer versioning documentation has been clarified.
MLflow 3.15.0 introduces an MCP Registry for registering and sharing Model Context Protocol servers, enhances the Assistant with multi-provider LLM support and per-session token usage tracking, and enables proxy-less artifact transfers via presigned URLs to reduce server load and timeouts on large files. Additional improvements include sharable Runs table views, multi-modal image attachments for LLM judges to evaluate vision tasks, and numerous bug fixes across tracing, evaluation, gateway, and UI components.
MLflow 3.14.0 adds one-command agent setup with Databricks support and durable low-latency Claude Code tracing, Review Queues for trace annotation and feedback collection, and @mlflow.test pytest markers for regression testing. Default model serialization formats change for sklearn to skops, PyTorch to pt2, and LightGBM to skops.
MLflow 3.13.0 introduces Role-Based Access Control with Admin UI, automatic trace archival to S3, and one-click observability for Claude Code and other coding agents. Breaking changes include a redesigned permission system (legacy APIs removed), MLServer removal from pyfunc serving, and requirement for MLFLOW_ALLOW_FILE_STORE=true flag for local file-based stores.