To view this email as a web page, click here

November 10th, 2026

Hello Builders. This month, we’re giving you more ways to move from experiments to better models faster. Explore new in-chat visualizations with ARIA, compare multimodal outputs with synced image controls, bring Optuna and Ax into Sweeps, and accelerate reinforcement learning with RL Rollouts. Plus, SDK 0.31.0 delivers faster startup, deeper system monitoring, and improved reliability across your training workflows. Check out the CoreWeave Forge documentation.

New: ARIA brings more visualizations directly into chat
 
ARIA can now render scatter plots as in-chat panels, with one point per run, making it easier to explore relationships between metrics and configuration values across experiments. New in-chat panel types also include parallel coordinates, parameter importance, run comparer, and media panels, alongside existing line, bar, and scatter plots. These expanded visualizations help Sweeps users, robotics teams, and other advanced users analyze experiments and uncover insights without leaving the conversation. Try ARIA for free now!
scatter_plots_demo.png

New: Use Optuna and Ax optimization engines with Sweeps

Teams can now use Optuna or Ax as their preferred optimization engine within Sweeps, bringing optimization, orchestration, experiment tracking, run queues, charts, and lineage into one workflow. Users can customize Optuna samplers and pruners or Ax generation strategies, define conditional search spaces, and optimize for multiple objectives such as accuracy and latency. This gives teams greater flexibility to use the optimization tools they prefer while keeping the full experimentation workflow connected in CoreWeave Forge. Learn more.

New: RL Rollouts accelerates reinforcement learning at scale

Now in private preview, RL Rollouts lets you keep your existing RL trainer while CoreWeave Inference serves the rollout layer at scale and hot-loads new checkpoints into live deployments with minimal downtime. In a joint validation with NVIDIA and You.com, RL Rollouts helped improve Nemotron 3.5 Lightning’s web-search accuracy from 36.97% to 45.45%, reduce tool calls by 30.24%, and load checkpoints 15 times faster than traditional redeployments. Learn more.

New: Compare multimodal outputs with synced image zoom and pan

You can now zoom and pan across multiple images simultaneously in W&B Models, making it easier to compare model outputs side by side. Zooming or moving within one image automatically keeps the others aligned at the same scale and position, while bounding boxes, masks, and class labels move with their images. With new zoom controls and shortcuts available in both panel and full-screen views, teams working on multimodal models, autonomous driving, and robotics can inspect visual differences without manually aligning each image. Get started with W&B Models.

Image (1).png

New: SDK 0.31.0 brings faster performance and deeper training visibility

SDK 0.31.0 is packed with performance, reliability, and monitoring improvements. CoreWeave Forge is now the default endpoint, with existing API keys continuing to work without requiring users to log in again. The release also makes SDK imports and CLI startup approximately twice as fast, improves offline run syncing, and adds new metrics for GPU energy consumption, CPU throttling, and out-of-memory events. It also introduces enhanced LEET debugging and monitoring capabilities, security updates, Python 3.15 support, and reliability fixes across run history, sweeps, integrations, and system metrics. View the full changelog.

Events & Webinars

Join the next What’s New Wednesday Webinar on November 18 to learn what CoreWeave announced at GTC Berlin. See how Physical AI Launchable Templates streamline robotics workflows such as driving scenario evaluation and synthetic scene generation. We’ll also cover Zero Trust AI with Self-Managed CrowdStrike and the upcoming NVIDIA HGX Rubin NVL8. Learn how tested configurations, standardized tooling, and customer-controlled security help teams build, secure, and scale complex AI workloads, and bring your technical questions for a live Q&A with our AI experts.

On November 24, join the Training Tuesday Webinar for a technical deep dive into turning experiment data into confident research decisions. See how CoreWeave Forge brings together experiment tracking in W&B Models, custom analysis and debugging in CoreWeave Notebooks, AI coding agents connected through MCP, and automated research workflows with CoreWeave ARIA. Through live demos, learn how to compare training runs, identify high-performing checkpoints, diagnose anomalies, and determine what to try next, so you can spend less time piecing together results and more time improving your models!


That's all for now.

The Weights & Biases Team