bison-1
Intro
bison-1 is a decision model that Baseten released on 20 November 2026 alongside groundhog-1. Both are built on Google’s Gemma 4, for teams that want an American base model. Official blog: Meet bison-1 and groundhog-1. The official account @baseten said the same day that the new models improve on mammoth-1 and elephant-1 in quality, keep similar speed, and add a 128k context window plus image-and-text input.
According to the official post, bison-1 is #1 of 19 decision models on Artificial Analysis Decision Index v2.1 with a score of 76, ahead of d1.2-ultra at 75 and mammoth-1 at 71. It also has the lowest Brier score on the index at 0.14 (lower is better), ahead of d1.2-ultra at 0.15. The models are live on Baseten Model APIs, with open weights on Hugging Face. The blog says bison-1 is priced like mammoth-1.
Advertisement
Explain
What problem it targets
Last month Baseten released mammoth-1 and elephant-1 on Qwen backbones. This launch aims for similar decision quality on Gemma 4, an American base model, with much longer context and multimodal input. The official example is an invoice that looks ready to pay, while an earlier email shows the order was canceled — without enough history, a model can approve a payment it should reject.
Official capability and training notes
- Context: 128k, eight times the previous decision models on Model APIs;
- Input: text and images;
- API: the same as the existing decision models; trying them is a model-name change;
- Calibration: the blog attributes part of the Brier-score gain to the Gemma base, and part to scaling “soft labels” in the new training mixture (for example “83% answer A, 17% answer B”) instead of treating the answer as 100% certain;
- Pareto front: the lab says bison-1, mammoth-1, and elephant-1 sit on the accuracy-versus-time Pareto frontier at their respective sizes.
Limits
bison-1 and groundhog-1 are two models launched in the same post; do not treat groundhog-1’s expected score as bison-1’s result. Decision Index and Brier figures are the official numbers on AA Decision Index v2.1. Exact Hugging Face repo names should be taken from Baseten or Hugging Face at the time you use them.
Usage
Who it is for
- Teams doing review, classification, policy, or document-style decisions who want a Gemma 4 backbone;
- Users of mammoth-1 or elephant-1 who want to compare the new quality and the 128k context window;
- Anyone who needs open weights to evaluate the model in their own environment.
How to use it
- Read the official blog: https://www.baseten.co/blog/meet-bison-1-and-groundhog-1 ;
- On Baseten Model APIs, change the model name to
bison-1(the blog says the API is the same as the existing decision models); - For self-hosting or inspection, follow the official pointer to the open weights on Hugging Face;
- When you quote the leaderboard, keep the official frame: AA Decision Index v2.1, bison-1 at 76, first of 19, Brier score 0.14.