mammoth-1 and elephant-1: Baseten Trains AI Models to Make Decisions Instead of Generating Text

Baseten releases mammoth-1 and elephant-1, two AI models specialized in classification and decision-making, with actionable confidence scores.

Not every artificial intelligence system needs to write an answer. Routing a customer request, detecting phishing, or determining whether content meets predefined rules can involve unnecessary computation when a language model is required to generate sentences. Baseten is taking a different approach with mammoth-1 and elephant-1, two decision models containing 27 billion and 9 billion parameters, respectively. Developed by its internal research team, Base Labs, the models select among predefined answers and assign probabilities to the available options. Introduced on October 21, 2026, they occupy leading positions in specialized decision-making benchmarks. Both are announced as open-source releases and are available to try through Baseten's API at no charge for one week. Models That Choose Instead of Writing

Large language models typically generate responses one token at a time. Even when an application simply needs a yes-or-no answer, it uses a system capable of producing long sequences of text. That flexibility is valuable for complex tasks, but it can create unnecessary latency and expense for applications that repeatedly make short, structured decisions.

Decision models avoid conventional text generation. They receive context, a question, and a set of possible answers, then return a probability distribution across those options. For example, a customer support request may be assigned probabilities corresponding to billing, security, or technical support. The application can choose the highest-scoring category or escalate the case when confidence is insufficient.

Baseten describes three supported operations: selecting among alternatives, scoring an input on a defined scale, and making binary decisions. A model can therefore classify a message, determine its urgency, or estimate whether an email resembles a phishing attempt. Its output is designed for software to process directly rather than for presentation as conversational prose.

The approach builds on the decision-model category associated with TypeSafe's Jev and System One API. It does not replace generative language models, but targets tasks where the available answers can be specified in advance. Confidence Scores and Human Escalation

The purpose of returning probabilities extends beyond choosing an answer. The scores allow applications to establish thresholds for automated action. A customer service system, for example, might automatically route a ticket when one category receives more than 90% confidence and send ambiguous cases for human review.

That distinction becomes particularly important in professional software. An incorrectly routed request may cause a delay, while a mistaken decision affecting security, payments, or account permissions can have more serious consequences. Overall accuracy is therefore not enough. Teams also need to check whether a model's stated confidence corresponds to how often its predictions are correct.

Baseten says decision models can be post-trained to improve this confidence calibration. Nevertheless, organizations must establish thresholds using representative data, including ambiguous cases and costly errors. A high model-generated probability should not automatically be treated as a reliable guarantee.

Decision models can also act as routers between other AI systems. Straightforward requests may be sent to smaller, less expensive models, while harder questions are escalated to frontier models. Baseten suggests decision models may be up to ten times faster than some reasoning-model approaches, although its launch announcement does not establish a universal latency improvement across tasks and deployment environments. mammoth-1 Leads the Jev Decision Index

Base Labs reports results from two separate leaderboards. On the Jev Decision Index, mammoth-1 scores 71.24 on version 0.4, placing it first among the 208 models included in the ranking cited by Baseten. Elephant-1 ranks lower overall, but is among the leading models in its nine-billion-parameter class.

On the separate Artificial Analysis Decision Index, mammoth-1 scores 65.9, while elephant-1 reaches 65.2. Baseten reports that the smaller model ties for first place with d1-ultra on that evaluation. These results should not be conflated with the Jev ranking, whose tasks and scoring methodology differ.

The evaluations cover knowledge and reasoning, language, tools and automation, retrieval and classification, and judgments involving artistic or human preferences. The aim is to test decision-making across varied domains rather than measuring only straightforward text classification.

A strong benchmark ranking, however, is not the same as proven reliability in production. Performance on general evaluations does not guarantee comparable results in a specialized support operation, regulated workflow, or cybersecurity application. Models must still be assessed using domain-specific data and the actual cost of incorrect decisions. Two Qwen3.5 Models Trained on 6.21 Million Examples

Neither mammoth-1 nor elephant-1 introduces an entirely new foundation architecture. Base Labs post-trained Qwen3.5-27B and Qwen3.5-9B for decision-making tasks. Both models were trained using 6.21 million examples combining public datasets and synthetically generated cases.

The training data spans core decision skills, safety and abstention, and specialized evaluation domains. Some examples specifically teach the models to acknowledge when available information is insufficient to make a reliable choice.

The ability to respond with “I don't know” is an important feature of the training approach. In automated applications, a confident but incorrect classification can be more damaging than an explicit refusal to decide. A decision model must therefore learn to distinguish cases where it has sufficient evidence from those requiring more information or human involvement.

Base Labs reports refining the training mixture through repeated training and evaluation cycles with elephant-1, before