Baseten joey-1, 2B scale with P50 20ms, $0.01 per API token

• Baseten released joey-1, a decision model based on Qwen3.5-2B, and provided weights on Hugging Face
• P50 latency is 20ms for 2,000-token requests, approximately 2.5 times faster than elephant-1 (49ms)
• Priced at $0.01 per 1M API input tokens, cheaper than elephant-1 ($0.035) and mammoth-1 ($0.10)
• Scored 63.30 on the public split of the Decision Index (ranked 1st among models under 4B), but the private Full score is low at 35.91

Performance and Benchmarks

• Based on the Jev Decision Index Full score, joey-1 (35.91) is lower than StartLux-Decision v1.1 2B (40.11)
• elephant-1 has a low AA-Omniscience accuracy of 11.6% but a high non-hallucination rate of 93.2%
• mammoth-1 recorded a JevBench Capability score of 85.2, ranking 1st on the open weights board

Technical Features and Architecture

• Returns probabilities for each option instead of generating sentences, processing requests as bundles of state and question
• Maintains the Gated DeltaNet structure of Qwen3.5 but changes only the full attention layers to block bidirectional masking
• Of the 6.21 million training data entries, 58.3% are core decision skills and 16.1% are safety and judgment abstention data
• Introducing judgment abstention training resulted in a 5.5-point increase in the Decision Index with a 4% increase in data