Baseten Mammoth-1 is a 27-billion-parameter AI model that never writes a sentence. You ask it a question, and it returns a probability for each possible answer. Then you decide what to do with that number.
On 21 October 2026, Baseten’s research team, Base Labs, released Mammoth-1 and its smaller sibling, Elephant-1, as open-weight models. On the same day, Vercel added both to its AI Gateway, free to use until 27 October 2026.
Key takeaways
- Mammoth-1 (27B) and Elephant-1 (9B) are “decision models”. They classify, route and score input instead of generating text.
- Both are fine-tuned from Alibaba’s Qwen3.5 models on 6.21 million decision examples. The weights are on Hugging Face.
- Baseten says Mammoth-1 ranks first of 208 decision models on the Jev Decision Index, with a score of 71.24.
- Both models were free on Baseten’s Model APIs and Vercel AI Gateway from 21 to 27 October 2026.
- The benchmark figures come from Baseten itself. No independent scores had been published at launch.
What is a decision model, and why not just use a chatbot?
A decision model answers typed questions about a piece of input, which Vercel calls the “state”. It returns a probability for every possible answer. It does not write prose.
There are three question types. Boolean asks a yes/no question and returns one probability. Choice picks one option from a named list, such as billing, shipping or technical. Score rates something on an ordered scale, such as low, medium or high urgency.
Think of an airport departures board. A chatbot is the gate agent who explains the delay in a friendly paragraph. A decision model is the board itself: one status per flight, with no small talk. If you need to sort ten thousand support tickets an hour, you want the board.
Baseten’s argument, set out in its launch post, is that using a general-purpose LLM for classification is slow and expensive. The text it produces also gives you no reliable sense of how sure it is. A decision model skips text generation entirely. Baseten claims this can be up to 10x faster than a reasoning model, though it gave no specific latency figures.
Baseten Mammoth-1 vs Elephant-1: how do they compare?
Both models share a 16,000-token context window, so each request can hold a long email thread or a short document. Here is how they differ, using Baseten’s own numbers.
| Detail | Mammoth-1 | Elephant-1 |
|---|---|---|
| Parameters | 27 billion | 9 billion |
| Base model | Qwen3.5-27B | Qwen3.5-9B |
| Jev Decision Index 0.4 | 71.24 (#1 of 208) | 62.6 (top 9B model) |
| Artificial Analysis Decision Index | 65.9 | 65.2 (tied #1 with d1-ultra) |
| Best suited to | Tasks needing broad knowledge | Fast, cheaper decisions |
| Gateway model ID | baseten/mammoth-1 | baseten/elephant-1 |
The odd result is Elephant-1. It is a third the size of Mammoth-1 but sits only 0.7 points behind it on the Artificial Analysis index. It ties there with Liquid AI’s d1-ultra, which we covered when Vercel added it to AI Gateway.
How were the models trained?
Base Labs fully fine-tuned the Qwen3.5 base models on 6.21 million decision-making examples. These mix public datasets with synthetic data. The training covers core decision skills, safety and abstention, and judging tasks in expert fields.
Abstention is the interesting part. Some examples teach the model to say it doesn’t know rather than guess. According to AICoder’s breakdown, the models use block-bidirectional attention and shared-prefix caching. In practice, that means one pass over the input can answer several questions at once. AICoder also reports that the weights carry an MIT license, citing the Hugging Face model pages.
How do you call Mammoth-1 on Vercel AI Gateway?
Per Vercel’s changelog, you have three routes:
- AI SDK: call
experimental_decidefrom theaipackage with a model ID, the state and a questions object. - OpenAI-compatible Decisions API: point the OpenAI client at Vercel’s gateway URL and call
decisions.create. - TypeSafe-compatible API: for teams already using TypeSafe’s Jev model, so they can switch models without rewriting their requests.
One request can ask several questions about the same message. Take “I was charged twice, please refund the duplicate.” You can ask whether it is a refund request, which team should handle it and how urgent it is, all in one call. Each answer comes back with its own probability.
Your code then sets a threshold. Baseten’s example routes a ticket automatically at 90% confidence and sends anything lower to a person. Vercel’s decision docs also describe opt-in fallbacks, which re-run a low-confidence answer on a second model. Both calls are billed when that happens.
Should you trust the benchmarks?
Not yet, and not blindly. Every number in this piece comes from Baseten or from articles repeating its claims. As RuntimeWire points out, how a model performs in real workflows matters more than where it sits on a leaderboard.
Our take: open weights are the real story here. You can run these models yourself, inspect them and measure them on your own hard cases. That matters more than a few points on an index. Before you set an auto-accept threshold, test how well calibrated the model is: when it says 90%, is it right about 90% of the time? Our AI agent evaluation checklist is a good place to start.
What happens next
The free window on both Baseten and Vercel ended on 27 October 2026. Baseten’s deadline was 11:59 PM Pacific time. After that, decision calls on AI Gateway are billed at each model’s per-token rates, which are listed on Vercel’s models page. Neither company published paid prices in its launch materials.
The decision-model category is getting crowded. TypeSafe’s Jev, Liquid AI’s d1 family, Microsoft’s Decision-1, Perplexity’s Decider and an OpenAI Decisions API all sit in the same Vercel lineup. Expect independent benchmark runs and price cuts to follow. Base Labs has also published a technical report, so outside researchers can try to reproduce its results.
FAQ
Are Mammoth-1 and Elephant-1 open source?
Yes. Both are open-weight models, published on Hugging Face under the baselabs account. AICoder reports an MIT license, citing the model pages.
Can Mammoth-1 write text or chat?
No. It returns probabilities for boolean, choice and score questions. That makes it a fit for routing, moderation and grading, not for writing answers.
Which one should I pick?
Baseten recommends Mammoth-1 for tasks that need broad general knowledge and Elephant-1 for speed and lower cost. The best way to choose is to test both on your own data.
Your next step: take 100 real tickets or messages your team has already labelled, run them through both models and compare the confidence scores with the right answers before you automate anything.
Leave a comment