Google Opens Vizier, the Tool That Tuned 70 Million Objectives Internally
Google's battle-tested black-box optimizer, powering 70M+ internal tuning jobs, is fully open-source and ready for your ML pipelines
Takeaways
- What it is: Google OSS Vizier is an open-source Python framework for black-box and hyperparameter optimization, based on Google's internal tuning service.
- Scale proven: The internal Google Vizier has tuned over 70 million objectives across Google products and research systems.
- Default algorithm: Uses a JAX-powered Gaussian Process Bandit with trust regions, output warping, and a vectorized Firefly acquisition optimizer.
- Broad problem support: Handles multi-metric, early stopping, transfer learning, conditional search, and mixed continuous/categorical spaces.
- Benchmarks: Outperforms Ax, HEBO, Optuna, and others on high-dimensional, categorical, batched, and multi-objective settings.
- Free and open: Apache 2.0 license; install with
pip install google-vizier[jax]; full documentation on ReadTheDocs.
Open Source Vizier is Google's production-grade black-box optimization framework, now available as a standalone Python package. It's the same system that has quietly powered hyperparameter tuning across Google's largest products and research efforts for years, and it's free to use under the Apache 2.0 license.
The problem it solves
Most optimization problems in ML aren't differentiable. You can't backprop through "which learning rate schedule works best" or "what neural architecture beats the baseline." These are black-box problems: you have inputs, you get outputs, and the function in between is opaque. Manual grid search wastes compute. Random search is better but dumb. What you really want is a system that learns from every trial and gets smarter about where to look next.
Blackbox optimization is the task of optimizing an objective function where the output is the only available information. Due to its generality, it has been applied to hyperparameter optimization, drug discovery, reinforcement learning, and industrial engineering. Vizier handles all of these under one unified API.
Battle-tested at Google scale
Google Vizier has performed millions of optimizations and accelerated numerous research and production systems at Google, demonstrating the success of Bayesian optimization as a large-scale service. The numbers are staggering: it has tuned more than 70 million objectives across Google, and is also available to the public via Vertex Vizier, a Google Cloud product.
Vizier is the de-facto blackbox and hyperparameter optimization service across Google, having optimized some of Google's largest products and research efforts. To operate at the scale of tuning thousands of users' critical systems, Google Vizier solved key design challenges in providing multiple different features, while remaining fully fault-tolerant.

What it can actually do
OSS Vizier provides an API capable of defining and solving a wide variety of optimization problems, including multi-metric, early stopping, transfer learning, and conditional search. In practice, this covers a lot of ground:
- Hyperparameter tuning: learning rates, batch sizes, architecture choices, regularization