Baseten releases AI models that score decisions instead of writing answers

Baseten released two AI models on October 21, 2026, for software that needs a decision rather than a written response. Built by its internal research team, Base Labs, Mammoth-1 and Elephant-1 return probabilities for supplied answers. The aim is to let applications act on confident choices and escalate uncertain ones instead of parsing generated prose.

Mammoth-1 has 27 billion parameters; Elephant-1 has 9 billion. Baseten says both are open-source on Hugging Face and available for a free one-week trial through Baseten Model APIs. The company also says the API offering has zero data retention enabled by default.

A choice, not a conversation

The distinction is the output. A conventional language model can classify a support ticket, but it generates text that software must interpret. These models instead receive an input and questions with defined answers. They return a probability for each option, with the probabilities adding up to one, rather than composing a reply.

Questions can ask for a choice from a list, a yes-or-no answer, or a rating on an ordered scale. That supports jobs such as routing tickets, detecting phishing, checking requests for safety, and selecting which language model should handle a request. The result is meant for the application’s next step, not a conversational exchange.

Baseten argues that skipping text generation makes these decisions faster and cheaper to serve. Confidence thresholds provide another control: its example routes a ticket automatically at 90% confidence and sends less-certain cases to a person. That is a configurable workflow example, not a demonstrated accuracy guarantee.

A claimed lead, with different testing contexts

Baseten says Mammoth-1 ranks first among 208 models on the Jev Decision Index 0.4, scoring 71.24. Elephant-1 ties for eighth overall and leads the 9-billion-parameter models. The index combines 37 public benchmarks with an undisclosed number of private tests, according to Base Labs’ technical report.

The comparison with rival d1-ultra needs a separate qualification. On Base Labs’ own replication of the Artificial Analysis Decision Index, Mammoth-1 scored 65.9 and Elephant-1 scored 65.2. The report compares those results with d1-ultra’s official score of 65. Those figures come from different testing contexts, rather than one shared official evaluation.

For choosing between its own models, Baseten recommends Mammoth-1 when a task needs more general knowledge and Elephant-1 for quick reasoning. Its evaluation advice goes beyond rankings: test difficult cases, check confidence, and measure latency and cost with retries and fallbacks included.

Changing what the model sees—and when it abstains

The technical report describes two main ingredients: adapting Qwen3.5’s attention and building a targeted training mixture. In full-attention layers, the researchers let tokens read their entire input or question block, rather than only preceding tokens. This lets a question be encoded with knowledge of its answer options. Other recurrent layers retain their original causal structure.

The design also reuses the processed input across questions. A support ticket, for example, is processed once; each additional question requires a processing pass over that question’s instructions and options alone. The input does not read the questions, so its stored computation remains reusable.

Both models were trained on a final mixture of 6.21 million examples. Base Labs developed that mixture through repeated experiments with the smaller model before training both sizes on it. The four categories show where the team spent its training budget:

• Core decision skills: 3.62 million examples spanning reasoning, language, tools, retrieval, classification and human taste.
• Safety and abstention: 1 million examples covering harmful requests, unsuitable tools and situations where the model should decline to answer.
• Judging and expert domains: 1 million examples involving rubrics, answer comparisons, and fields including law, finance and health.
• Public decision data: 590,000 examples drawn from public datasets, reasoning probes and community decision datasets.

The abstention work addressed a specific weakness: Base Labs says its earlier experiments struggled to recognize when not to decide. The added examples included harder multiple-choice questions where a wrong answer was worse than admitting uncertainty, making refusal part of the training objective rather than treating every question as requiring a choice.