Baseten releases mammoth-1 and elephant-1, open-source decision models

Rankings and availability

Baseten says mammoth-1 leads the Jev Decision Index 0.4 with a score of 71.24, and elephant-1 is the top 9B model on that index. The company adds that elephant-1 is tied for #8 overall. On the separate Artificial Analysis Decision Index, mammoth-1 scores 65.9 and elephant-1 scores 65.2. Baseten states that elephant-1 is tied for #1 there with the previous best model, d1-ultra.

Both models are open-source on Hugging Face. They can also be tried free for a week on Baseten Model APIs, with zero data retention (ZDR) by default. Baseten built them by post-training Qwen3.5-9B and Qwen3.5-27B. It points to TypeSafe's introduction of Jev and the System One API as the start of decision models as a new category.

What decision models do

Unlike generative LLMs, decision models do not generate text. They return a probability distribution over the possible answers to a question: a score for every option, with all scores summing to 1. Baseten describes three question types:

• Choice: pick an option from a list, such as which team owns a ticket.
• Score: rate the input on a scale, such as whether a ticket is low or high urgency.
• Boolean: answer yes or no, such as whether an email is a phishing attempt.

According to Baseten, using generative LLMs for classification tasks like routing a ticket or flagging spam is costly and slow, because the computation is overkill for the output required. Their text response also gives no reliable measure of certainty. Decision models skip text generation, which Baseten says makes them cheaper and faster to serve in production. Compared with a reasoning model, a decision model may give up to a 10x speed up, the company says.

Confidence thresholds can be customized. A support ticket, for example, can be routed automatically when the model is 90% confident and sent to a person when confidence is lower. The models can also be post-trained so their confidence better reflects how often they are right. Listed use cases include ticket triage and routing, spam and phishing detection, prompt guardrails and jailbreak detection, and routing requests between small and frontier LLMs.

Training and model choice

Baseten built the models as generalists for tasks ranging from routing support tickets to checking requests for safety. Training and evaluation used data from five domains: knowledge and reasoning, language, tools and automation, retrieval and classification, and arts and human taste.

Both models were fine-tuned on 6.21 million decision-making examples. The dataset combined public datasets with synthetically generated examples covering core decision skills, safety and abstention, and judging and expert domains. This included examples that taught the model to answer "I don't know" when it was uncertain about the accuracy of its answer. Details are in a technical report from Baseten's research lab, Base Labs.

Baseten recommends mammoth-1 for tasks that need more general knowledge, and elephant-1 for those that mainly need quick reasoning.

What it means for teams evaluating decision models

Baseten notes that the two models are part of a growing set of decision models, and suggests teams test candidates against their own workflow. It lists three criteria: accuracy, including difficult cases; confidence, meaning whether the model knows when it is unsure, so a threshold can be set for accepting or escalating answers; and speed and cost, measured as latency and cost per decision at expected traffic, including retries and fallbacks.

With open weights on Hugging Face and a free week on Baseten Model APIs, teams can run such comparisons without an upfront commitment. The rankings are Baseten's own reporting of results on third-party indexes.