Both are open-source models from Base Labs, Baseten's internal research team, built by post-training Qwen3.5-27B and Qwen3.5-9B respectively.
Rather than generating text, the models return probability distributions over supplied answers for choice, scoring, and Boolean tasks. Baseten names ticket routing, phishing and spam detection, prompt guardrails, and routing between language models as intended uses. The company says it trained the models on 6.21 million decision-making examples.
Baseten reports mammoth-1 scored 71.24 and ranked first of 208 models on Jev Decision Index 0.4; it calls elephant-1 the leading 9B model. Both are available through Baseten Model APIs with a one-week free trial and zero data retention by default, and their weights are available on Hugging Face.
The announcement does not establish independent reproduction of the benchmark results, or production latency, cost, calibration, and accuracy for a particular deployment.