Is Weights & Biases (wandb) Down? How to Check wandb Status in 2026

Weights & Biases (wandb) is the experiment tracking platform trusted by ML teams at OpenAI, NVIDIA, Toyota Research, and thousands of other organizations. It's often the first tool researchers reach for when a training run starts — which means when wandb goes down, it can stall entire ML pipelines.

Whether you're seeing wandb: ERROR Failed to sync messages, a hanging wandb.init() call, or an empty dashboard that should have new runs, this guide will help you determine: is wandb down, or is there something wrong on your end?

How to Check if wandb is Down (Fastest Methods)

1. Check the Official wandb Status Page

Weights & Biases maintains a live status page at status.wandb.ai. It shows real-time uptime for the API, artifact storage, dashboard, and any ongoing incidents. If there's a confirmed outage, it will be posted here with timeline updates.

2. Test wandb Connectivity from Your Terminal

Run a quick connectivity check to see if the wandb client can reach the API:

# Check wandb status
wandb status

# Test API reachability
python -c "import wandb; wandb.login()"

# Quick API health check
curl -I https://api.wandb.ai/healthz

A ConnectionError or timeout confirms the wandb API is unreachable. A 401 response means an auth issue with your API key, not an outage.

3. Check Your Training Script for Hanging wandb.init()

If wandb.init() hangs indefinitely when starting a new run, wandb is likely unreachable. You'll see this in training scripts as a silent hang rather than an explicit error. Check your process and kill it — then switch to offline mode (see below).

4. Search X (Twitter) for ML Community Reports

Search "wandb down" or "Weights Biases outage" on X filtered by Latest. The ML community reports tool outages on X very quickly — often within the first few minutes of an incident starting.

5. Use API Status Check for Automated Monitoring

API Status Check monitors wandb's API endpoints continuously and sends instant alerts via Slack, email, or PagerDuty — so you can switch to offline mode before a long training run gets stuck mid-epoch.

How to Keep ML Training Running When wandb is Down

Unlike most SaaS tools, wandb has an excellent offline mode that lets your training continue uninterrupted even during a full platform outage. Here's how to use it:

Switch to Offline Mode Before Starting Training

# Set offline mode via environment variable
export WANDB_MODE=offline

# Or set it in your Python script
import wandb
import os
os.environ["WANDB_MODE"] = "offline"
wandb.init(project="my-project", name="experiment-1")

# Or pass it to init directly
wandb.init(project="my-project", mode="offline")

In offline mode, wandb writes all metrics, logs, and artifacts to a local directory (./wandb/). Your training script runs exactly as normal — you just won't see live updates in the cloud dashboard.

Sync Offline Data After wandb Recovers

# Sync all offline runs in the current directory
wandb sync ./wandb

# Sync a specific run
wandb sync ./wandb/offline-run-20260630_072900-abc123

# List all pending offline runs
wandb sync --list

Once the wandb platform is back up, run wandb sync to upload all locally buffered run data. Your experiment history will appear in the dashboard as if there was no outage.

Why Does wandb Go Down?

Weights & Biases handles massive amounts of data — metrics from thousands of simultaneous training runs, large model artifacts, dataset versions, and rich media logs. This creates specific failure patterns:

• Artifact Storage Overload: wandb stores model checkpoints and datasets in cloud storage (GCS/S3). A surge in large artifact uploads — common after a major model release or research deadline — can overwhelm the storage pipeline, causing artifact sync failures and API timeouts.
• Metrics Ingestion Backpressure: When thousands of runs submit metrics simultaneously (end-of-quarter research sprints, competition deadlines), the time-series ingestion pipeline can back up, causing dashboard lag and API slowdowns.
• Dashboard Frontend Failures: wandb's rich charting and visualization layer occasionally has separate outages from the API backend. You may be able to log data while the dashboard is unavailable, or vice versa.
• Authentication Service Issues: wandb's auth layer handles API key validation for millions of runs. An auth service hiccup causes widespread 401 Unauthorized errors even with valid keys.
• Planned Maintenance: Weights & Biases schedules maintenance windows for infrastructure upgrades. Check their status page and changelog for advance notices.
• GCP/AWS Regional Issues: wandb's infrastructure runs on GCP and AWS. Regional cloud provider incidents can cause localized wandb failures for users in affected geographies.

Common wandb Errors and What They Mean

What to Do When wandb Is Down

1. Switch to offline mode immediately: Set WANDB_MODE=offline in your environment. Restart any training scripts that were hanging waiting for the API. Your runs will continue and buffer locally.
2. Check status.wandb.ai: Confirm it's a platform-wide outage, not a local issue. Subscribe to incident updates for real-time recovery notifications.
3. Verify your API key is still valid: Run wandb login --relogin to refresh your credentials. Sometimes key expiration coincides with apparent outages.
4. Continue training without interruption: Offline mode is fully functional — all metrics, logs, system stats, and artifact references will be tracked locally and uploaded on sync.
5. Sync when recovered: Run wandb sync ./wandb after the platform is back up. Your full experiment history will appear in the dashboard.
6. Set up automated monitoring: Use API Status Check to monitor the wandb API endpoint. Get an alert the moment wandb goes down — and again when it recovers.

wandb Alternatives for ML Experiment Tracking

If you need experiment tracking while wandb is down or want to diversify your ML observability stack:

• MLflow: Open-source experiment tracking with self-hosted or managed options. Works without any external connectivity — runs a local server you control.
• Comet ML: Cloud-based experiment tracking with similar feature set to wandb, including artifact versioning and team collaboration.
• TensorBoard: Google's built-in visualization toolkit for TensorFlow and PyTorch. Fully local, no API dependency, and zero latency.
• Neptune.ai: Hosted experiment tracking platform with strong metadata management and large artifact support.
• Aim: Open-source experiment tracker with a local-first architecture. All data stays on your machine until you choose to share it.

Frequently Asked Questions