Base Labs announced joey-1 on October 29, 2026, as a small decision model for high-volume routing, classification, and scoring. Its 2B-class positioning and hosted pricing make it a candidate for narrow gates inside larger agent workflows. The announcement date, checkpoint upload history, and hosted availability are separate milestones.
The upload came before the announcement
The primary records establish this sequence:
- October 27, 2026: the Hugging Face history records an initial commit and an “Upload joey-1” commit. This establishes a repository upload, not when the repository first became publicly accessible.
- October 29, 2026: model-card updates added a banner and the launch-blog link. Baseten’s dated announcement said joey-1 was available on Model APIs and with open weights.
- November 1, 2026: these dates distinguish the milestones rather than assigning all availability to one launch date.
The October 29 announcement describes availability, not a waitlist or preview. It confirms API access by that date; it does not establish the first moment the endpoint accepted requests. Meanwhile, the repository contains an actual model.safetensors file, not merely a promise of future weights. Neither record proves a separate earlier preview.
A decision endpoint, not a text-writing assistant
Baseten’s library documents text and image inputs, structured decision outputs and the /v1/systemone endpoint. Its example uses the model identifier decision-models/baseten-joey-2b-latest. Questions can request a choice, a yes/no assessment or an ordered score, with probabilities rather than free-form explanations. Published Model API pricing lists $0.01 per million input tokens. These are vendor specifications, not results from our own tests.
For an operations team, the useful distinction is between choosing a destination and executing an action. A ticket can be classified for billing without giving the classifier permission to issue a refund. Our assessment: joey-1’s most plausible role is that narrow decision boundary, with a separate workflow enforcing permissions and handling exceptions.
What the speed and accuracy evidence supports
Baseten reports 20-millisecond P50 end-to-end latency for an average 2,000-token request at production scale, versus 49 milliseconds for elephant-1. This is a vendor-reported median under its serving conditions—not a latency guarantee for a customer’s network, image workload or traffic peaks.
The model card reports 63.30 on the Jev Decision Index 0.4 public split, compared with elephant-1’s 72.24, and says joey-1 trails that larger sibling in every evaluated area. Its Artificial Analysis Decision Index result is explicitly an internal replica, not an independently certified leaderboard result. The release therefore supports investigating a speed–accuracy tradeoff, not assuming a small model can replace every decision system.
Operationally, evaluate the expensive mistakes separately: a wrongly routed ticket and an incorrectly cleared security request should not share one acceptance threshold. Probability outputs are useful inputs to escalation policies, but do not make those policies self-validating.
Downloadable weights still have deployment boundaries
The MIT-labeled repository provides self-hosting instructions requiring one CUDA GPU and a pinned serving runtime. Its model card specifies a 16,384-token maximum sequence length and says overlength requests are refused rather than truncated. Baseten separately advertises 16k on Model APIs and 262k on dedicated deployments. Do not treat that dedicated-deployment claim as proof that the downloadable recipe has been validated at the larger length.
Our recommendation for teams assessing this release is a shadow-routing trial: replay representative cases without allowing actions, measure wrong-route and escalation rates, and inspect peak latency alongside median latency. Preserve the tested weights revision or document the hosted model identifier and evaluation date. The announcement expands deployment choices; it does not remove the need to prove that a faster decision is an acceptable one.