Decision models skip text generation: mammoth-1 tops 208 models, elephant-1 leads the 9B class

Decision models form a separate category: they do not write an answer, but return a probability distribution summing to 1 over all candidate answers, handling choice from a list, scoring on a scale, and yes/no questions. Using generative LLMs for classification tasks such as ticket routing or spam flagging is costly and slow because their computation far exceeds the required output, and the response alone gives no reliable measure of certainty. mammoth-1 ranks first among 208 decision models on the Jev Decision Index 0.4; elephant-1 is tied for eighth on that index and is the leading 9B decision model, and is tied for first with the previous best model d1-ultra on the Artificial Analysis Decision Index.

Both models were post-trained from Qwen3.5-9B and Qwen3.5-27B on a dataset mixing public datasets with synthetically generated examples covering core decision skills, safety and abstention, and judging and expert domains. The data mix was refined through repeated training and evaluation runs with elephant-1, and the final mix was then used to train both models. Training used 6.21 million decision-making examples, including samples teaching the model to answer "I don't know" when uncertain; evaluation covered five domains: knowledge and reasoning, language, tools and automation, retrieval and classification, and arts and human taste.

This report covers the model specifications, training-data scale and rankings on two decision indices as published by the releasing lab. It is aimed at engineers who need classification, routing or safety decisions in production; the material recommends evaluating accuracy, confidence, and speed and cost, and setting a confidence threshold to decide when to accept an answer or escalate. Evaluation details and per-item scores behind the rankings are not described.

The material states decision models may give up to a 10x speedup over reasoning models but gives no measurement conditions or baseline configuration. Rankings on both indices are reported as ties, and per-item scores are not listed. The material says post-training can make confidence better reflect how often the model is right, but reports no measured calibration error.