Baseten's Base Labs releases mammoth-1 and elephant-1 open decision models

What Base Labs released

Base Labs, the internal research team at the AI inference company Baseten, released two open decision models on 21 October 2026: mammoth-1, with 27 billion parameters, and elephant-1, with 9 billion. Baseten's announcement calls them two of the world's best decision models, and both sets of weights are on Hugging Face as baselabs/mammoth-1 and baselabs/elephant-1.

Both are full fine-tunes of Alibaba's Qwen3.5 models, Qwen3.5-27B and Qwen3.5-9B. Base Labs says mammoth-1 ranks first of 208 models on the Jev Decision Index 0.4 with a score of 71.24, and that elephant-1 is the strongest model under 15 billion parameters on that index. The accompanying technical report is by Max Kirkby, Michael Psenka, Thomas Villeneuve, Aaryam Sharma and Charles O'Neill.

Decision models are a category that took off after TypeSafe introduced Jev and its System One API in September 2026; the company has since raised a large Series A. Since then Microsoft, Cloudflare, Liquid AI, Perplexity and others have released their own, and Base Labs is now entering with open weights under the MIT licence.

What a decision model does

A decision model does not write text. It takes a state, such as a support ticket or a document passed as JSON, plus one or more questions, and returns a probability for each allowed answer in a single forward pass. Base Labs describes three question types: a choice among 1 to 255 named options, a yes-or-no question that returns one probability, and a score on a scale of 2 to 10 levels.

Baseten's pitch is cost and control. Because nothing is generated, there is no output to parse and the work per request is small; the company says a decision model can be up to ten times faster than a reasoning model. The probabilities let developers set thresholds, for example routing a ticket automatically when the model is 90 percent sure and sending it to a person otherwise.

Suggested uses include ticket triage and routing, spam and phishing detection, prompt guardrails and jailbreak detection, and deciding whether a request should go to a small model or a frontier one. It is the same pitch made for Microsoft's Decision-1 and Cloudflare's Clef, which mammoth-1 now claims to beat on the Jev index.

How the models were built

Qwen3.5-27B is a hybrid model. Each of its 16 repeating units has three Gated DeltaNet layers, a recurrent type of layer that only looks backward, followed by one full-attention layer. Base Labs changed only those 16 full-attention layers, replacing the usual causal mask with a block-bidirectional one, and left the DeltaNet layers causal. elephant-1 uses the same design on Qwen3.5-9B, with eight units and 32 layers against mammoth-1's 64.

The change splits each request into a shared prefix, holding the system prompt and the state, and one suffix per question. Inside a block, tokens can see each other in both directions, so a question is read with knowledge of its options. The prefix never looks at the questions, so its cache can be reused for every question in a request. In a single ablation run on an early data mix, the bidirectional mask scored 48.6 against 44.8 for the causal one on a public sample of the Jev index. The report notes that Perplexity's Decider v1.1 makes a similar change.

Training was one epoch over 6.21 million examples with the vision encoder frozen. The loss combines cross-entropy with a Brier term that penalizes probability placed on wrong answers, which Base Labs says discourages overconfidence. elephant-1 took 5,546 steps on 128 H100 GPUs and mammoth-1 took 9,859 steps on 72 GB300 GPUs. The team says it did the work in about a week.

The data recipe behind the scores

Base Labs credits most of its gains to data. The final mix, the eighth it tried, contains 3.62 million examples of core decision skills, 1 million on safety and abstention, 1 million on judging and expert domains such as law, finance and health, and 0.59 million converted from public datasets.

Synthetic examples were generated with models such as DeepSeek V4-Pro or GLM-5.3, then labelled by a panel of five judge models and kept only if at least four agreed, or labelled exactly by a program where possible. The pipeline then removes ambiguous items and any overlap with the benchmarks, has a separate agent look for shortcuts, and tests each new source in a 240-step pilot training run.

On the team's own runs of the public Jev suite, the 9B model rose from 45.5 with the first 254,000-row mix to 66.0 at 4.19 million rows. Most of the remaining gain came from fixing data rather than adding it: an audit of all 253 sources for wrong labels and giveaways helped take the score to 73.9, and the final mix reached 74.0.

The models were also trained to say "I don't know". When a model's top probability fell below a threshold a little under 0.5, the training data taught it to choose an explicit uncertainty option, or to pick it when the correct answer had been removed from the list.

Benchmark results, as reported

On the Jev Decision Index 0.4, whose full score the report says is computed by the board's maintainers on private tests, mammoth-1 scores 71.24. That is 1.8 points ahead of Perplexity Decider v1.2 at 69.42, 9.0 points ahead of Jev 1.14 and 15.5 points ahead of Cloudflare Clef. The private tests are 60 percent the same skills and 40 percent new domains.

elephant-1 scores 62.65 on the same index. The report calls that tied seventh overall, while Baseten's blog says tied eighth; both place it first among models under 15 billion parameters. Base Labs also worked with Artificial Analysis, whose decision index gives elephant-1 a 65, level with Liquid AI's d1-ultra. On Base Labs' own replica of that index, mammoth-1 scores 65.9 and elephant-1 65.2.

The report adds that on the Jev index's private tests of the same skills, both models are narrowly beaten in three of five areas, even though mammoth-1 leads four of five areas on the public suite.

How to try them

Both models are free to try for a week on Baseten Model APIs, with zero data retention by default, through a decision endpoint at /v1/systemone. The Hugging Face repositories include the weights, a serving bundle and setup scripts. The model cards set the licence field to "other" with the licence name MIT, and Baseten's model library lists MIT.

Baseten suggests evaluating any decision model on accuracy, including hard cases, on whether its confidence matches how often it is right, and on latency and cost per decision at real traffic levels. Teams already fine-tuning Qwen3.5 for classification, such as Coinbase with its fraud model, now have a ready-made open starting point to compare against.

What to keep in mind

Every score here is reported by Base Labs or Baseten. "Best in the world" claims rest on two leaderboards in a category that is weeks old, where the leader has changed several times, including when Perplexity open-sourced Decider v1.2. AI Lookout has not run the models.

The two Baseten sources disagree slightly on elephant-1's overall rank. The bidirectional-attention ablation is a single run. Base Labs says it decontaminated its data against the benchmarks it reports, but it used models to generate and label most of the training set, and its own decision benchmark has not been released yet.

The models keep Qwen3.5's vision encoder but were trained on text only, so image and video inputs are untested; multimodal training, "looped" models that spend more compute on hard decisions and a fully automated data loop are listed as future work.