Rollout · NOV 2026Bison-1 and Groundhog-1: frontier American-backbone decision models

Meet bison-1 and groundhog-1, our new decision models built on Gemma 4, with 128k context, multimodal support, and open weights.

chuck

Charles O'Neill

maxkirkby

Max Kirkby

thomas

Thomas Villeneuve

Nataly Delcid

Nataly Delcid

Michael Psenka

Michael Psenka

We’re expanding our decision model family with bison-1 and groundhog-1, built on Google’s Gemma 4. They bring our work on decision models to an American backbone and longer context, now up to 128k tokens.

Our full announcement on the Baseten blog covers the release and benchmark results. A few highlights:

  • Number 1 on the Artificial Analysis decision index. At launch, bison-1 ranks first among 19 models on Artificial Analysis Decision Index v2.1, scoring 76. It also records the index’s lowest Brier score, a measure of how closely confidence tracks correctness.

  • More context for each decision. Both new models support 128k context, eight times our previous Model API offering, and accept text and images.

  • Two model sizes. bison-1 is a 31B dense model focused on accuracy; groundhog-1 is a 26B mixture-of-experts model with roughly 4B active parameters, designed for lower latency and cost.

  • Open weights. Both are available on Hugging Face and Baseten Model APIs, using the same API as our existing decision models.

Read the full announcement here.

Cite this work

@article{baselabs2026-bison-1-and-groundhog-1-frontier-american-backbone-decision-models,
  title   = {Bison-1 and Groundhog-1: frontier American-backbone decision models},
  author  = {O'Neill, C. and Kirkby, M. and Villeneuve, T. and Delcid, N. and Psenka, M.},
  journal = {Base Labs},
  year    = {2026},
  number  = {009},
  url     = {https://labs.baseten.co/009},
}

Related from the lab

01

Can gradient alignment make supervision on-policy?

Rollout·NOV 2026Michael Psenkachuckmaxkirkby
02

The bias of on-policy self-distillation

Rollout·NOV 2026maxkirkbyMichael Psenkachuck
03

Understanding the compute efficiency of value functions in RL

Rollout·OCT 2026HarryAly ShariffmaxkirkbyMichael Psenka