Baseten released bison-1 and groundhog-1 on November 20, 2026, two open-weight models post-trained from Google’s Gemma 4 to return calibrated probabilities for predefined decisions. Artificial Analysis scored bison-1 at 76 on Decision Index v2.1, placing it first among 19 evaluated decision models.
Both are available through Baseten Model APIs and as weights on Hugging Face. They accept text and images and support 131,072 tokens; Baseten says that is eight times the 16K context of its previous mammoth-1, elephant-1 and joey-1 APIs. A request can bundle one state and any number of yes/no, named-choice or ordered-score questions into one shared-prefix prefill, without generating text.
Two Model Sizes, Shared Deployment Terms
Bison-1 is a dense 31B-parameter model post-trained from Gemma 4 31B. Groundhog-1 is a 26B mixture-of-experts model based on Gemma 4 26B-A4B, with about 4B active parameters per token. Baseten prices them at $0.10 and $0.035 per million tokens, respectively, matching mammoth-1 and elephant-1.
Both use the same API as Baseten’s earlier decision models, and requests above the trained length limits are rejected rather than truncated. Google’s Gemma 4 base models carry Apache License 2.0 terms; Baseten’s modifications are released under the MIT License, with use subject to both licenses.
Independent Decision Index Results
The Artificial Analysis Decision Index v2.1 combines six evaluations conducted independently by Artificial Analysis. AA-Paperclip, covering decisions and extraction across 1,500 resumes, receipts, invoices and research papers, carries 30% of the weight. The When2Call tool-calling evaluation carries 10%, while tests of workplace judgement, knowledge and hallucination, reasoning and decision invariance each carry 15%.
Its Decision Score subtracts a Brier calibration penalty from accuracy. The underlying probabilities also support confidence thresholds that answer some cases and route others elsewhere, making calibration relevant to how a decision system is deployed.
Artificial Analysis’s November 19, 2026 results place d1.2-ultra second at 75 and mammoth-1 at 71, while bison-1’s 0.14 Brier score is the index’s lowest, against 0.15 for d1.2-ultra. The Brier measure calculates the mean squared gap between the probability assigned to the chosen answer and whether that answer is correct. Lower values indicate confidence that more closely matches actual outcomes.
The bison-1 model card reports first-place results on When2Call, EvalSafe-O*NET-AA, HLE-MultiChoice v1.1 and AA-Symmetry v1.1, plus second place on AA-Paperclip. Its 60 on AA-Omniscience-Decision v1.1 ranked 11th, so the aggregate lead includes substantial variation across components rather than uniform dominance.
Groundhog-1’s model card states that it has not yet received an Artificial Analysis Decision Index score.
Training Data and Long-Context Tests
Both models were trained on 6.71 million decision examples spanning core decision skills, safety and abstention, judging and expert domains, documents and long-context tasks, including public benchmark sets. The objective combines cross-entropy with a Brier calibration term and includes a trained I don't know option. About one example in seven carries a probability target instead of a single right answer.
Baseten attributes a significant calibration gain to expanding these soft-label targets alongside the Gemma base. Training used 192 B200 GPUs for bison-1 and 128 for groundhog-1.
In Baseten’s own long-context runs at 128K tokens, bison-1 scored 84.8 on LongMemEval M and 87.1 on LongMemEval S, while groundhog-1 scored 77.4 and 79.3, respectively. These company-run tests address long-context performance separately from Artificial Analysis’s broader six-part index.
The release extends long-context, multimodal decision making across both of Base Labs’ established price tiers. Groundhog-1’s independent index result remains the open question for judging whether its mixture-of-experts design delivers the mammoth-1-level performance Baseten expects.




