Flow

MLflow is great for local experimentation but self-hosting it for a team can be painful. Weights & Biases is a hosted product with no equivalent of the simple mlflow ui workflow. Flow combines the strengths of both: lumlflow ui is local and easy to start, and experiments can be uploaded to LUML when you're ready, which brings multi-user collaboration and the rest of the platform (registry, deployments, monitoring, and more) on top of experiment tracking.

No. Flow works entirely offline: run pip install lumlflow, run your script, then run lumlflow ui. You can use Flow indefinitely without ever signing up. A LUML account is only needed when you want to upload experiments to a shared workspace.

Call tracker.enable_tracing() to route OpenTelemetry spans into the experiment store, then activate the relevant instrumentor before the agent is built. For example, use instrument_openai() for OpenAI clients or LangchainInstrumentor().instrument() for LangChain or LangGraph. Any traced call made inside the experiment becomes a span.

For ML: static parameters via log_static, step metrics via log_dynamic, model artifacts via log_model, and arbitrary file attachments. For GenAI: full trace trees with inputs, outputs, and timings, eval samples with scorer results, and human annotations. Tags and experiment groups apply to both.

Locally, in a SQLite database. By default that's sqlite://./experiments, the directory you run lumlflow ui from. Override it with the --path flag or the BACKEND_STORE_URI env var. When you upload an experiment, the model and its context go to the LUML registry. LUML is managed, but the storage and compute it orchestrates can live wherever your team runs them.

The log_model helper handles scikit-learn, XGBoost, LightGBM, CatBoost, and LangGraph with auto-detection. Other ML frameworks like PyTorch, TensorFlow, JAX, and Keras work through the same SDK by calling log_static, log_dynamic, and log_model directly. LLM clients (OpenAI, Anthropic, LangChain, LlamaIndex, DSPy) are traced via OpenTelemetry instrumentors.

Yes. Once an experiment is uploaded, anyone with access to your LUML workspace sees the same parameters, metrics, traces, and artifacts you do. The local UI also supports cross-experiment comparisons within a group, useful before you decide what to push.

An uploaded experiment becomes a versioned artifact in the LUML registry. From there you can deploy it to a Satellite, monitor it, or use it as a baseline elsewhere on the platform.