A 2B decision model that returns calibrated probabilities in about 20 ms, built for high-volume routing and classification.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Why joey-1 Answers in 20 ms: No Text Generation, One Forward Pass
Benchmarks
joey-1 on the Jev Decision Index: Best Under 4B, Behind Its Siblings
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
joey-1 or elephant-1: Speed and Cost Versus Decision Accuracy
joey-1 and elephant-1 are the 2B and 9B members of Baseten's Base Labs decision-model family. Both are trained on the same data mixture and share the same System One API, so Baseten describes switching between them as a one-line change. The practical decision is a trade-off. Baseten reports 20 ms P50 for joey-1 versus 49 ms for elephant-1, and joey-1 is priced lower. On the Jev Decision Index 0.4 public split, however, elephant-1 scores 72.24 against joey-1's 63.30.
Choose joey-1 when decision volume is high, latency is added to every user-facing response, and a first-pass answer with escalation for uncertain cases is acceptable. Choose elephant-1 when decisions carry more risk, need broader general knowledge, or when joey-1's accuracy on your labeled data falls short. Baseten positions elephant-1 as the balanced default. A common pattern is joey-1 first, with elephant-1 handling low-confidence cases.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Support ticket triage and routing
Baseten lists ticket triage and routing as a core decision-model use case. joey-1 can answer several typed questions about one message in a single request, such as team, urgency and refund eligibility, with the questions sharing an input prefill. A 20 ms reported median latency and low per-token cost suit very high ticket volumes. Tickets that come back with low confidence can be escalated to elephant-1 or mammoth-1.
Guardrails and abuse screening
Baseten cites spam and phishing detection and prompt-guardrail or jailbreak checks as decision-model use cases. A guardrail sits in front of every request, so its latency adds to every response. That is where joey-1's speed matters most. Training included a safety-and-abstention data share and a trained 'I don't know' option, which helps surface uncertain inputs. Borderline cases still warrant a larger model or human review.
First-pass gate in LLM cascades
Baseten recommends joey-1 as a first-line model, with elephant-1 or mammoth-1 as the escalation path when it is uncertain. It also lists routing between small and frontier LLMs as a use case. Because joey-1 returns a probability, you can set a confidence threshold: act on confident answers and pass the rest on. This keeps expensive generative calls for requests that need them.
Rubric scoring and yes/no checks
joey-1 supports ordered scoring levels as well as yes/no and multiple-choice questions, so it can score content against a rubric and return typed answers your code can use directly. It suits bulk quality checks and content classification. It does not explain a score, and it is less accurate than its larger siblings on harder judging tasks.