Baseten announced on November 20 the bison-1 and groundhog-1 models, two new open-weight decision models built on Google's Gemma 4. Bison-1 has taken the top spot on the Artificial Analysis Decision Index, the benchmark for this kind of model, with a score of 76 out of 19 models evaluated — ahead of Liquid AI's d1.2-ultra (75) and Baseten's own mammoth-1 (71).
Decision models are not general-purpose chatbots: they take a question with defined options and return calibrated probabilities for each possible decision — approving or rejecting a payment, classifying a document, choosing an agent's next action. That is why calibration matters as much as accuracy: bison-1 also posted the lowest Brier score on the index, 0.14, meaning the confidence it states is the closest to how often it is actually right.
What changes from the previous generation
Both models offer a 128k-token context window (131,072), eight times the 16k of their predecessors mammoth-1, elephant-1 and joey-1, and accept text and image input. In practice, that allows full policies or complete conversation histories to be included in a decision, without summarizing or cutting excerpts. Baseten gives the example of an invoice that looks ready to pay, but whose approval changes when an earlier email shows the order was cancelled.
Bison-1 is a dense 31-billion-parameter model, trained from Gemma 4 31B. Groundhog-1 is a 26-billion-parameter mixture-of-experts (MoE) model, with about 4 billion active parameters per token, based on Gemma 4 26B-A4B. Pricing matches the equivalent predecessors: $0.10 per million tokens for bison-1 and $0.035 per million for groundhog-1.
Open weights and an American backbone
Both models are already available on Baseten Model APIs, with the same API as the earlier models — switching models means changing the name in the call — and with open weights on Hugging Face. The Gemma 4 base carries the Apache 2.0 license, and Baseten's modifications are released under the MIT License. According to the company, the models were trained on 6.71 million decision examples.
Baseten presents the pair as the first in the family on what it calls an American backbone: mammoth-1, elephant-1 and joey-1 were built on Qwen models. Geekiki already covered the launch of joey-1, the smallest in the family, with 2 billion parameters: Baseten launches joey-1, a 2B decision model at $0.01 per million tokens. The company says groundhog-1 does not yet have an official score on the Artificial Analysis index, but it expects the model to match mammoth-1's performance.
Image: official bison-1 artwork by Baseten / Base Labs (Baseten's official blog).
Sources: Baseten — Meet bison-1 and groundhog-1; MarkTechPost; Unite.AI.
Transparency note: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources on the internet, and reviewed before publication. Check the original sources to confirm the information.
Por Muse