Baseten Releases joey-1, an Open-Weight 2B Decision Model Priced at $0.01 per Million Tokens
The 2B model returns calibrated probabilities in a single forward pass, and its MIT-licensed weights are now on Hugging Face.
Baseten has released joey-1, a two-billion-parameter decision model built by Base Labs on top of Qwen3.5-2B, together with its weights under the MIT license. Unlike a chat model, joey-1 takes a state — text or image — plus typed questions such as multiple choice, yes/no, or sequential scoring, and returns calibrated probabilities for every option in a single forward pass, with no free-form text to parse. It was trained on 6.21 million decision examples using a cross-entropy objective plus a Brier calibration term, and ships with a trained "I don't know" option.
On the Jev Decision Index 0.4 the model scores 63.30, placing it first among models under 4B parameters and 37th of 230 models overall. Median latency is about 20 milliseconds for 2,000-token requests, roughly 2.5 times faster than Base Labs' larger elephant-1 (9B), and API pricing is $0.01 per million tokens.
joey-1 anchors a decision-model family — joey-1 (2B), elephant-1 (9B), mammoth-1 (27B) — in which low-confidence answers escalate to a bigger model or a human reviewer, an architecture aimed at high-volume classification and routing workloads.