JOURNAL / 2026.10.26

Mammoth-1 and Elephant-1 offer decisions with probabilities and open weights

Baseten releases two models specialized in choosing, classifying and scoring. Their usefulness depends on more than accuracy: deciding what to automate and what to send for review.

A message says: “You charged me twice and I need my money back.” A support system might need three answers: which team should handle it, whether it requests a refund and how urgent it is. Writing a full explanation to extract those answers afterwards adds work the application may not need.

Baseten has introduced Mammoth-1 and Elephant-1, models with 27 billion and 9 billion parameters specialized in decisions. They return probabilities over options, yes-or-no answers and scores. They do not invent automated classification: they offer another way to handle varied tasks through instructions, without generating a text response that must then be interpreted.

Availability is concrete. Vercel added them to AI Gateway on October 21, with temporary free access through October 27. The longer-term development is downloadable weights: the Mammoth-1 model card lists an MIT license and describes self-hosting with a CUDA GPU and a specific environment. Publishing weights lets users retain and adapt the model; it does not mean it runs on any computer. Its documented window is 16,384 tokens.

The Base Labs technical report explains why this involves more than a JSON format. Starting from Qwen3.5, they modify full attention so that each block can be read in both directions. The document is processed once; each question reuses that information and adds computation for its options. Answers come from internal scores for those options, without developing a text sequence. This could be particularly useful when one document needs many classifications.

Specialization also limits the appropriate assignment. Choosing between known teams fits this interface. Investigating why a charge was duplicated, consulting transactions and negotiating a resolution requires a broader workflow. My interest is in inserting a bounded decision where we currently use an entire conversation, provided evaluation shows that the substitution pays off.

There are capability signals, although their provenance matters. The card reports that Mammoth-1 outperforms Perplexity Decider v1.2 on Jev Decision Index 0.4. The report describes private tests of known skills and new domains. For Artificial Analysis, it distinguishes an Elephant-1 evaluation conducted with that evaluator from its own index replicas. This is evidence for considering the models, rather than turning positions on different leaderboards into a customer-support guarantee. We have not measured their latency or cost per decision either.

The most delicate question arises with automation. A distribution summing to one can be confidently wrong. The report distinguishes an option’s probability from the confidence field, which rescales the concentration of that distribution; these values are not interchangeable. Training against overconfidence helps, but does not establish calibration in another language or for a different class of incidents.

I would test on historical messages reviewed and held out for evaluation. Among decisions assigned high probabilities, I would check how many are correct and which errors remain. I would separate languages and rare classes. Then I would choose the threshold according to the harm of being wrong: sending a case to another queue and authorizing a refund require different criteria. This is an evaluation proposal, not an observed result for these models.

Vercel’s decision documentation already supports configuring a second query when the first returns uncertainty. This is optional and involves two-stage billing. It provides the escalation mechanism; the team must establish whether the second answer improves the outcome and what it costs.

This looks like a useful route towards smaller, more inspectable automations. The benefit will come from defining options, measuring acceptable errors and preserving a way out for uncertain cases. Calling a prediction a “decision” does not supply those properties. Open weights make it possible to revisit that choice with more independence from the provider.

Primary sources consulted on October 26, 2026. We reviewed announcements, documentation and published artifacts; we have not run the models or reproduced their evaluations. The opening message is a conceptual example.

← Back to journal