Baseten released joey-1 on Thursday, a 2B-parameter decision model from its in-house Base Labs team, priced at $0.01 per million tokens and available as open weights on Hugging Face.
A decision model does not generate text. Ask it a question and it hands back a label with a confidence attached, in a single forward pass, and the scores across every candidate sum to one. Feed joey-1 some text, optionally with images, plus a set of questions, and it scores every option at once by sharing one prefix through the model. Yes or no, a pick from named choices and graded levels are all supported.
It joins mammoth-1 at 27B and elephant-1 at 9B, both released October 21. The three were built on one recipe and answer through one interface, so a team can swap models without rewriting code. Mammoth-1 ranked first among 208 models on the Jev Decision Index at that release, and elephant-1 the best at 9B. The positioning is volume. Buyers who cannot afford to be wrong reach for mammoth-1, elephant-1 covers the middle, and joey-1 exists for sheer throughput where milliseconds count.
The economics are the interesting claim. Consider a help desk triaging 10 million tickets monthly, each around 500 tokens, for a total near 5 billion. Baseten bills that workload at roughly $50 a month on joey-1, and ten times as much on mammoth-1. Latency follows the same pattern. At production scale a 2,000-token request returns in a 20 ms median, against 49 ms for elephant-1. The company describes joey-1 as a first-line model and the larger siblings as the escalation path for hard cases.
Reported numbers need care. The model card puts joey-1 at 63.30 on the Jev public split, 37th of 230, well behind elephant-1 at 72.24.
