Baseten Open-Sources a Model That Never Writes a Word

On October 21, Base Labs — the internal research arm of AI infrastructure company Baseten — published open weights for two models that deliberately don't do what most open-source releases this year have been racing to do. Mammoth-1 and Elephant-1 don't write prose, hold a conversation, or generate code. Given a situation and a set of possible answers, they return a calibrated probability for each one and stop. Both are published on Hugging Face under the baselabs organization with an MIT license, and both are fine-tunes of Alibaba's Qwen3.5 family: Mammoth-1 starts from Qwen3.5-27B, Elephant-1 from Qwen3.5-9B.

A model built to decide, not to talk

Base Labs calls these "decision models." Instead of generating a response token by token, a decision model takes a text state — optionally with images — plus a list of candidate answers, and returns a probability distribution across three question formats: choice (pick one from a list), score (rate on a scale), and boolean (yes or no). Base Labs pitches the approach at tasks that generative chat models handle clumsily and expensively: support-ticket routing, spam and phishing detection, prompt guardrails that catch bad inputs before they reach a more expensive model, and routing requests between smaller and frontier-scale models. The company's own benchmarking claims the approach runs up to 10 times faster than sending the same classification task to a reasoning model, since there's no token-by-token generation step to wait on.

The training recipe behind both models — what Base Labs labels "Mix 8" — ran 6.21 million decision-making examples through full fine-tunes of the Qwen3.5 checkpoints (the vision encoder was kept frozen). According to the Hugging Face model card, 58.3% of that mix targets core decision skills, 16.1% is devoted to safety and abstention — explicitly teaching the model when to answer "I don't know" rather than guess — another 16.1% covers judging and expert-domain questions, and the remaining 9.5% comes from public decision datasets. The team credited on the release includes Max Kirkby, Michael Psenka, Thomas Villeneuve, Aaryam Sharma, Charles O'Neill, Chloe Florit, and Ardin Lo; none of them is quoted directly in the announcement.

Benchmark claims, and a telling distribution strategy

Base Labs is leaning on two outside scoreboards to make its case. On the Jev Decision Index (version 0.4), it says Mammoth-1 ranks first out of 208 listed decision models with a score of 71.24, while the smaller Elephant-1 is the top-ranked 9B-parameter model and ties for eighth overall. On the separate Artificial Analysis Decision Index, Mammoth-1 scores 65.9 and Elephant-1 scores 65.2, tying Elephant-1 for first place alongside a model called d1-ultra. Those are Base Labs' own citations of third-party leaderboards rather than independently reproduced numbers, and as RuntimeWire's coverage of the release noted, the real test is whether the leaderboard gains translate into production workflows rather than just benchmark position.

The go-to-market move is as notable as the models themselves. Baseten is giving away access to both models on its own Model APIs for one week with zero data retention by default, and it simultaneously pushed Mammoth-1 and Elephant-1 onto Vercel's AI Gateway, where Vercel's own changelog confirms they're free to call — for input and output alike — through October 27, with a 16,000-token context window and usage logged against normal AI Gateway budgets. Vercel's model listing describes the pair as "open-weights" but directs usage terms to Baseten's own policies rather than naming a license outright, a small inconsistency against the unambiguous MIT grant stated on the Hugging Face card itself.

What the release underscores is a quieter trend running alongside the race for bigger open frontier models: companies treating an existing open-weight base — in this case Alibaba's Qwen3.5 — as raw material for narrow, commercially packaged derivatives rather than chasing general-purpose chat leaderboards directly. Base Labs didn't train a foundation model from scratch; it took an open one already competitive on general benchmarks and spent its compute budget teaching it a single, narrower skill — calibrated judgment — then shipped the result back out under a permissive license with a Baseten-hosted API as the on-ramp. Whether "decision models" become a durable category of open-source tooling or a one-off experiment will depend on whether classification and routing workloads actually move off generative LLMs once the free trial period on Baseten and Vercel ends next week.