LoginSign Up
← Back to all sparks
B

Baseten

AI-ASSISTANTS
Velocity5.0

AI model deployment and inference platform for running ML models in production.

www.baseten.co ↗

Baseten adds open-weight joey-1 and finer control over rolling promotions.

ai-inferencemodel-apisdecision-modelsdeployment-controlsobservability
◆Current state
Baseten added joey-1, a 2B-parameter decision model post-trained from Qwen3.5-2B, to Model APIs with open weights on Hugging Face. Rolling promotions now accept max surge and max unavailable together, so teams can set extra replica capacity and allowed serving loss in the same rollout. Before that came paid usage for elephant-1 and mammoth-1, and per-replica metrics.
◆Where it's heading
Model APIs are collecting small routing and classification models, now three in ten days, while dedicated deployments gain production controls: GPU metrics, per-replica breakdowns and now rollout tuning. The security tightening of late October appears to have settled into defaults.
◆Prediction
Expect more small decision models on Model APIs, possibly on the same free-then-paid pattern as elephant-1, and further deployment and rollout controls for dedicated inference.

◆Recent moves

  1. 14d ago

    Mixed-mode rolling promotions

    Rolling promotions can now combine max surge and max unavailable, so a rollout can trade extra capacity against temporary serving loss. It extends the production-control work on dedicated deployments from observing replicas to deciding how they change.

    View source ↗
  2. 15d ago

    joey-1 decision model available on Baseten

    joey-1, a 2B-parameter decision model post-trained from Qwen3.5-2B, joins Model APIs with open weights. It is the third classification and routing model in ten days, which makes small decision models a deliberate line on the platform.

    View source ↗
  3. 16d ago

    elephant-1 and mammoth-1 pricing

    The two classification models added on October 21 leave their free launch window and become paid from October 27. It confirms the free-then-paid pattern Baseten uses for new Model APIs; the entry gives no price detail.

    View source ↗
  4. 16d ago

    Per-replica metrics

    Metrics can now be split per replica to compare CPU, GPU and memory and spot outliers hidden by averages. It extends last week's GPU temperature and utilization metrics into finer-grained debugging for dedicated deployments.

    View source ↗
  5. 20d ago

    Revocation of unused API keys on 10/24

    From October 24, keys unused for 30 days are revoked automatically. It replaces the 90-day November plan announced a day earlier with a stricter, immediate one, so teams with idle production keys need to check them now.

    View source ↗
  6. 21d ago

    Auth errors switch from 403 to 401 on October 28

    From October 28, missing or invalid credentials return 401 instead of 403. It is a small but breaking change for clients that branch on the status code.

    View source ↗