Mammoth-1 and Elephant-1 Decision Models Review as Base Labs Claims the Top Spot

Base Labs, a research lab inside Baseten, released two open-weight decision models, mammoth-1 (27B parameters) and elephant-1 (9B parameters), post-trained from Qwen3.5 and published on Hugging Face. Base Labs reports mammoth-1 scores 71.24 on the Jev Decision Index, first of 208 models and 1.8 points ahead of Perplexity Decider v1.2, while elephant-1 scores 62.65 and is the best model under 15B parameters; on the Artificial Analysis Decision Index Base Labs reports elephant-1 at about 65, level with Liquid AI's d1-ultra. The gains came mainly from data — a 6.21 million example training mix plus a training trick that teaches the models to say "I don't know" — and while the Jev leaderboard independently confirms the scores, regular API pricing and production speed figures remain unpublished.

A new research lab inside Baseten has released two open-weight models it calls the best decision models in the world. The claim is bold and specific. So it is worth looking at what the models are, how they were built, and how much of the evidence is independent. - Base Labs, a research lab within Baseten, released mammoth-1 27B and elephant-1 9B , decision models built by post-training Qwen3.5 models, with open weights on Hugging Face - A decision model reads a state and a set of typed questions and returns calibrated probabilities instead of text, a format TypeSafe introduced in September 2026 with Jev - Base Labs says mammoth-1 scores 71.24 on the Jev Decision Index, first of 208 models and 1.8 points ahead of Perplexity Decider v1.2, and elephant-1 scores 62.65 and is the best under 15B parameters - On the Artificial Analysis Decision Index it reports elephant-1 at about 65, level with the previous best, Liquid AI's d1-ultra - The big gains came from data: a 6.21 million example training mix, plus a training trick that teaches the models to say "I don't know" - The Jev leaderboard independently confirms the models' scores, while regular API pricing and production speed figures remain unpublished What a decision model is A decision model is not a chatbot. It takes a piece of state, such as a support ticket or a document, plus a list of questions, and returns a probability for each possible answer. The questions are typed: a choice among up to 255 options, a yes or no, or a score on a scale of two to ten levels. Nothing is generated, so there is no text to parse and no way for the model to ramble. TypeSafe introduced the format in September 2026 with Jev and its System One API, and we have explained the idea for founders before https://startupfortune.com/system-one-models-explained-what-are-they-and-why-should-founders-care/ . The appeal is practical. Because every answer is a probability, a developer can set a threshold: act automatically when the model is confident, hand off to a person when it is not. Base Labs argues this is also why the format is cheap. The cost is one pass over the input, and each extra question adds only a small forward pass over the question itself. Mammoth-1 and elephant-1 The two models are named after animals and sized accordingly. Elephant-1 has 9 billion parameters and mammoth-1 has 27 billion. Both are post-trained from Alibaba-family Qwen3.5 models, which the Base Labs post describes by size without naming the developer, and both ship with open weights on Hugging Face. Base Labs says it trained them over about a week. The main architectural change is small and specific. Qwen3.5 mixes linear-attention layers with full-attention layers. Base Labs altered the attention mask on only the full-attention layers, so that the state and the question can see each other in both directions: each question is read with knowledge of its options. A side benefit is that the cached encoding of the state can be shared across every question in a request. In a single-run test on the 9B model, the bidirectional version scored 48.6 against 44.8 for the standard causal one - which Base Labs itself flags as one run. Nace.AI's Drex v2.0 bets agents don't need words to make a decision https://startupfortune.com/naceais-drex-v20-bets-agents-dont-need-words-to-make-a-decision/ Nace.AI's Dos Bakytzhan built the Palo Alto startup on $26.5 million in funding from Jerry Yang and an Intel executive, and Drex v2.0 now edges rival Jev 1.13.0 by 3.3 points on the Decision Index benchmark. - AI agents decision making without language models https://startupfortune.com/naceais-drex-v20-bets-agents-dont-need-words-to-make-a-decision/ - Drex v2.0 probability scoring for agent decisions https://startupfortune.com/naceais-drex-v20-bets-agents-dont-need-words-to-make-a-decision/ The data did most of the work The post is candid that data mattered more than architecture. Its final training mix has 6.21 million examples, built over eight attempts. Most are core decision skills across knowledge, language, tools, retrieval and taste, with the rest covering safety and abstention, expert domains and public data from dozens of datasets. A panel of five judge models checked synthetic labels, and a label was kept only if at least four of the five agreed with the intended answer. Base Labs reports that elephant-1's public score climbed from 45.5 on the first mix to 74.0 on the last. The second ingredient is calibration. The training randomly adds an "I don't know" option and, when the correct answer is removed or the model is not confident enough, rewards it for choosing it. A squared-error term in the loss further discourages overconfidence. Earlier decision models, Base Labs says, "were terrible at knowing when they should not make a decision." It reports no standard calibration metric, so how well this works is not yet measurable from the post alone. The benchmark claims On the Jev Decision Index, a leaderboard of decision models, Base Labs reports 71.24 for mammoth-1, ranking it first of 208 models. That is 1.8 points ahead of Perplexity Decider v1.2 at 69.42. It is 9.0 points ahead of Jev 1.14 and 15.5 ahead of Cloudflare's Clef. Elephant-1 scores 62.65, tied for seventh overall and, in the company's words, the best decision model under 15 billion parameters. On the Artificial Analysis Decision Index, it puts elephant-1 at roughly 65, level with Liquid AI's d1-ultra, the previous best. These are strong numbers, and the full Jev scores are computed by the index's maintainers on private tests. The current leaderboard confirms mammoth-1 at 71.24 and elephant-1 at 62.65. Base Labs also says it worked with Artificial Analysis on elephant-1's pre-release evaluation, although its materials report that result as both 65 and 65.2. The technical report has a similar inconsistency in elephant-1's public Jev score, giving 72.2 in the text and 74.0 in a table. What is missing The technical report gives no latency or throughput results, and regular API pricing has not been published. The weights are available under the MIT licence, while Baseten offers both models through a hosted System One API, free through October 27, 2026. The models were trained on text only, with Qwen3.5's vision encoder left frozen, and Base Labs lists multimodal decisions and "looped" decision models as future work. It also says the decision benchmark it built itself will be released later. Why this matters to builders The release shows a fast-moving corner of AI that is turning into a crowded race. Jev, Perplexity's Decider, Cloudflare's Clef, Microsoft's model and now an open-weight pair from Baseten are all competing at the same narrow task. Open weights are the interesting part: a team can run elephant-1 on its own hardware for routing, grading and screening, instead of paying per call. That is the same job we recently moved to a small general model. A specialist that is both open and small is worth testing against it. Test it on your own questions before trusting any leaderboard, including this one. Recursive says AI built most of its new Flint coding model https://startupfortune.com/recursive-says-ai-built-most-of-its-new-flint-coding-model/ Recursive's Flint runs on an Apache 2.0 license with a 262,000-token context window and leads its size class on SWE-Bench Pro, SWE-Bench Verified, and Terminal-Bench. Those benchmark numbers come from Recursive itself, not an independent lab. - how to build AI with automated research https://startupfortune.com/recursive-says-ai-built-most-of-its-new-flint-coding-model/ - AI coding model built by machine learning https://startupfortune.com/recursive-says-ai-built-most-of-its-new-flint-coding-model/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Tagged for reading: Mammoth-1, Elephant-1, Decision Models, Review, Comparison, Base Labs, Top Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.