Baseten 开源决策模型 mammoth-1（27B）和 elephant-1（9B）

Baseten 工程博客·2026-10-22 00:58· 1 小时前AI 评分60

The top ranked decision models in the world: mammoth-1 and elephant-1

AI 导读

Baseten 研究实验室 Base Labs 发布开源决策模型 mammoth-1（27B）和 elephant-1（9B），基于 Qwen3.5-27B 和 Qwen3.5-9B 在 621 万决策样本上后训练而成。

正文 · 原文

We’re proud to release two of the world’s best decision models, trained by our own research lab: mammoth-1 (27B) and elephant-1 (9B)! 

mammoth-1 ranks #1 out of 208 decision models, and elephant-1 is the leading 9B decision model on the Jev Decision Index. Both models are open-source and free to try for a week on Baseten Model APIs, with zero data retention (ZDR) by default.

✕

mammoth-1 leads the Jev Decision Index 0.4 with a score of 71.24; elephant-1 is the top 9B model. On the Artificial Analysis Decision Index, mammoth-1 scores 65.9 and elephant-1 65.2.

Since TypeSafe introduced Jev and the System One API, decision models have emerged as a new category. We post-trained Qwen3.5-9B and Qwen3.5-27B to build two of the best. Here’s a closer look at their capabilities, use cases, and how to evaluate them for your stack.

What are decision models? 

Unlike generative LLMs, decision models don’t generate text. Instead, they return a probability distribution (a score for every possible answer, with all scores summing to 1) over the possible answers to each question.

You can ask decision models 3 types of questions: 

• Choice: pick an option from a list (e.g., which team owns this ticket: infrastructure, billing, security, or support?)

• Score: rate the input on a given scale (a 2 point scale would be low vs high). For example, how urgent is this ticket, low or high?

• Boolean: answer yes or no (e.g., is this email a phishing attempt?) 

Where do decision models beat generative LLMs?

Using generative LLMs for classification tasks, like routing a ticket or flagging spam, is costly and slow because the amount of computation they perform is overkill for the required output of the task. The response alone also doesn’t provide a reliable measure of certainty.

Decision models like elephant-1 and mammoth-1 skip text generation, which makes it cheaper and faster to serve in production. Customizable confidence thresholds help determine when to accept an answer and when to escalate it. For example, a support ticket 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’re right. 

In short, decision models are built to understand the request and make a decision without spending time writing a response. Compared to a reasoning model, a decision model may give up to a 10x speed up. 

Here are a few use cases:

• Ticket triage and routing

• Spam and phishing detection

• Prompt guardrails and jailbreak detection

• Routing requests between small and frontier LLMs

✕

How a decision model works. The input and output of a decision model when faced with a choice question. Here the state (a support ticket) and the question about the ticket (where to route it) go through one forward pass of the elephant-1 decision model and a probability for each option is returned.

What elephant-1 and mammoth-1 are built for

We built elephant-1 and mammoth-1 to be generalist decision models that could handle a broad range of tasks, from routing support tickets to checking requests for safety. We trained and evaluated decision-making using data from five domains, mirroring previous work: 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, while elephant-1 is tied for #8 overall and is the leading 9B decision model. On the separate Artificial Analysis Decision Index, elephant-1 is tied for #1 with the previous best model, d1-ultra. 

mammoth is a better fit if your task needs more general knowledge. If you mainly need quick reasoning, then choose elephant-1.

Architecture and training 

We fine-tuned Qwen3.5-9B and Qwen3.5-27B on 6.21 million decision-making examples. To train each model, we used a dataset that combined public datasets with synthetically generated examples covering core decision skills, safety and abstention, and judging and expert domains. We refined the data through repeated training and evaluation runs with elephant-1, then used the final mix to train both models.⁠ This included examples that taught it when to answer “I don’t know,” when the model was uncertain about the accuracy of its answer.⁠ Read more about how we trained it in our technical report here.

How to choose the right decision model

elephant-1 and mammoth-1 are part of a growing set of decision models. To choose the right decision model for your workflow, we recommend you evaluate: 

• Accuracy: Does it make the right decisions? Include difficult cases.

• Confidence: Does it know when it’s unsure? Check when to accept an answer and when to escalate, and set your confidence threshold accordingly. 

• Speed and cost: Measure latency and cost per decision at your expected traffic level. Include retries and fallbacks.

Try elephant-1 and mammoth-1 today

You can try elephant-1 and mammoth-1 for free for a week through Baseten Model APIs. Both models are also open-source on Hugging Face. You can read the technical report from our research lab, Base Labs, here. We look forward to seeing how you use them!

来源：Baseten 工程博客 · baseten.co