Base Labs opens mammoth-1 and elephant-1 under MIT, free on Baseten and Vercel for a week

Source · Model card: mammoth-1 (27B) + elephant-1 (9B)

On 21 October 2026 Baseten announced two decision models trained by its research team, Base Labs: mammoth-1 (27B, Qwen3.5-27B) and elephant-1 (9B, Qwen3.5-9B). The day before, both Hugging Face repositories were gated “for Decision Index evaluation only” (our 20 Oct note). They are now ungated and under the MIT license, with a technical report that covers the data mix (6.21M examples), the block-bidirectional attention mask and the Brier calibration term.

Where to run them. Baseten serves both on POST /v1/systemone (model page), and Vercel AI Gateway lists them as baseten/mammoth-1 and baseten/elephant-1 through the AI SDK, the Decisions API and a TypeSafe-compatible route. Both are free until 27 October, 11:59 PM PT; the price after that is not published. Self-hosting uses the bundled setup.sh on one CUDA GPU.

Board check. The Decision Index moved to edition 0.4.1 overnight and rescored mammoth-1 on its current public weights: Full 68.44, tied #3 (it was 71.24 and #1 in 0.4 on an earlier revision). elephant-1 is at 60.26 (#40), and Jev 1.13 at 59.22. On Artificial Analysis, Elephant 9B is 64.7, second to d1-ultra (65.2). Baseten and Vercel describe that as a tie for #1; we report the board’s numbers. mammoth-1 is not on AA, and neither model is on JevBench yet.

Catalog page: mammoth-1 + elephant-1.