Baseten Releases Two Decision Models

Baseten released mammoth-1 (27B) and elephant-1 (9B), open-source decision models trained by its internal research team, Base Labs, and made them free to try for a week through Baseten Model APIs. Zero data retention (ZDR) is enabled by default. Baseten says mammoth-1 ranks #1 among 208 decision models and leads the Jev Decision Index 0.4 with 71.24; elephant-1 is the leading 9B decision model. On the Artificial Analysis Decision Index, mammoth-1 scores 65.9 and elephant-1 65.2.

Decision models return a probability distribution across possible answers, with scores summing to 1, instead of generating text. They handle choice questions, ratings on a scale, and yes-or-no questions. Baseten says generative LLMs can be slow and costly for classification tasks such as ticket routing and spam detection, and their text responses do not directly provide a reliable confidence measure.

Baseten says skipping text generation makes decision models faster and cheaper to serve in production. Users can set confidence thresholds to accept answers or escalate uncertain cases to a person; for example, a support ticket could be routed automatically at 90% confidence. The company says post-training can make confidence better reflect accuracy, and decision models may be up to 10x faster than reasoning models. Suggested uses include ticket triage and routing, spam and phishing detection, prompt guardrails and jailbreak detection, and routing requests between small and frontier LLMs.

Baseten trained the models as generalists for tasks including support-ticket routing and safety checks. Their training and evaluation covered five domains: knowledge and reasoning, language, tools and automation, retrieval and classification, and arts and human taste. mammoth-1 ranks #1 on the Jev Decision Index 0.4; elephant-1 is tied for #8 overall and leads among 9B decision models. On the separate Artificial Analysis Decision Index, elephant-1 is tied for #1 with d1-ultra. Baseten says mammoth-1 is better suited to tasks needing more general knowledge, while elephant-1 is intended for quick reasoning.

The models were fine-tuned from Qwen3.5-9B and Qwen3.5-27B using 6.21 million decision-making examples. The dataset combined public data with synthetic examples on core decision skills, safety and abstention, and judging and expert domains. Baseten says it refined the data through repeated training and evaluation with elephant-1 before using the final mix to train both models. Examples also taught the models when to answer “I don’t know” if uncertain. The models are open-source on Hugging Face and available to try free for a week through Baseten Model APIs. Baseten recommends evaluating accuracy on difficult cases, confidence and escalation thresholds, and speed and cost at expected traffic levels, including retries and fallbacks.

Baseten released mammoth-1 (27B) and elephant-1 (9B), open-source decision models trained by its internal research team, Base Labs, and made them free to try for a week through Baseten Model APIs. Zero data retention (ZDR) is enabled by default. Baseten says mammoth-1 ranks #1 among 208 decision models and leads the Jev Decision Index 0.4 with 71.24; elephant-1 is the leading 9B decision model. On the Artificial Analysis Decision Index, mammoth-1 scores 65.9 and elephant-1 65.2.

Decision models return a probability distribution across possible answers, with scores summing to 1, instead of generating text. They handle choice questions, ratings on a scale, and yes-or-no questions. Baseten says generative LLMs can be slow and costly for classification tasks such as ticket routing and spam detection, and their text responses do not directly provide a reliable confidence measure.

Baseten says skipping text generation makes decision models faster and cheaper to serve in production. Users can set confidence thresholds to accept answers or escalate uncertain cases to a person; for example, a support ticket could be routed automatically at 90% confidence. The company says post-training can make confidence better reflect accuracy, and decision models may be up to 10x faster than reasoning models. Suggested uses include ticket triage and routing, spam and phishing detection, prompt guardrails and jailbreak detection, and routing requests between small and frontier LLMs.

Baseten trained the models as generalists for tasks including support-ticket routing and safety checks. Their training and evaluation covered five domains: knowledge and reasoning, language, tools and automation, retrieval and classification, and arts and human taste. mammoth-1 ranks #1 on the Jev Decision Index 0.4; elephant-1 is tied for #8 overall and leads among 9B decision models. On the separate Artificial Analysis Decision Index, elephant-1 is tied for #1 with d1-ultra. Baseten says mammoth-1 is better suited to tasks needing more general knowledge, while elephant-1 is intended for quick reasoning.

The models were fine-tuned from Qwen3.5-9B and Qwen3.5-27B using 6.21 million decision-making examples. The dataset combined public data with synthetic examples on core decision skills, safety and abstention, and judging and expert domains. Baseten says it refined the data through repeated training and evaluation with elephant-1 before using the final mix to train both models. Examples also taught the models when to answer “I don’t know” if uncertain. The models are open-source on Hugging Face and available to try free for a week through Baseten Model APIs. Baseten recommends evaluating accuracy on difficult cases, confidence and escalation thresholds, and speed and cost at expected traffic levels, including retries and fallbacks.