Baseten Release Notes

You can now gradually shift traffic to new deployments instead of swapping all at once. Candidate replicas scale up incrementally while previous replicas scale down in controlled steps, giving you zero-downtime updates. Pause, resume, or cancel mid-rollout if you spot issues. Configure rollout speed with max_surge_percent and stabilization_time_seconds, and choose how to handle the previous deployment after promotion.

Enable rolling deployments in your environment's promotion settings in the dashboard, or through the API:

curl -X PATCH "https://api.baseten.co/v1/models/{model_id}/environments/{env_name}" \
-H "Authorization: Api-Key $BASETEN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
  "promotion_settings": {
    "rolling_deploy": true,
    "promotion_cleanup_strategy": "SCALE_TO_ZERO",
    "rolling_deploy_config": {
      "max_surge_percent": 10,
      "stabilization_time_seconds": 300
    }
  }
}'

For more information, see Rolling deployments.

You can now terminate a specific replica within a deployment using the new management API endpoint. This lets you remove individual replicas without affecting the rest of the deployment, making it useful for evicting stuck or unhealthy replicas.

curl --request DELETE \
--url https://api.baseten.co/v1/models/{model_id}/deployments/{deployment_id}/replicas/{replica_id} \
--header
"Authorization: Api-Key
$BASETEN_API_KEY
"

For more information, see the terminate deployment replica docs.

We've redesigned the logs and metrics views for better visibility and faster debugging.

Logs

Logs now include an interactive volume chart that visualizes log frequency over time. Click a bar to zoom into that time window, or drag across a range to filter: the log viewer updates in real time. The logs view is now full-width by default, replacing the previous fullscreen toggle.

Metrics

Metrics also get the full-width treatment, along with a new grid view that displays charts in a two-column layout so you can see more metrics at once. A toggle at the top of the page lets you switch between grid and list views, and your preference is saved across sessions.

The Kimi K2 0905, Kimi K2 Thinking, and DeepSeek v3.2 Model API(s) were deprecated at 5pm PT March 6th. The model ID is currently inactive and will return an error for all requests.

As open source models advance rapidly, we prioritize serving the highest quality models and deprecate models when stronger alternatives are available.

We recommend Kimi K2.5 as an alternative. Kimi K2.5 offers very strong agentic coding and tool calling capabilities. Just swap in the new Model ID moonshotai/Kimi-K2.5

If you’d like to continue using the previous weights, please contact us about a dedicated deployment of the model.

We just launched the Baseten Delivery Network (BDN), designed to make cold starts 2-3x faster for large models.

BDN solves three root causes of slow cold starts: slow weight pulls from upstream storage, replica stampedes under load, and upstream availability dependencies. On first deployment, BDN mirrors your weights to secure storage. From there, a multi-tier cache (node cluster mirrored origin) serves weights with consistent hashing and single-flight semantics: each file fetched once per cluster, not once per pod. Fine-tunes sharing weights with a base model only pull the delta.

Check out the launch blog to learn more, or see the docs to get started.