Editorial illustration for Baseten's Bison-1 Model Tops Multiple Decision-Model Index Benchmarks
Baseten's Bison-1 Dominates Decision-Model Benchmarks
Baseten pushed two new models onto Hugging Face this week, and neither one writes a sentence of prose. bison-1 and groundhog-1 are decision models, built by post-training Google's Gemma 4 to spit out calibrated probabilities instead of text. Ask one a yes/no question, a multiple choice question, or a scoring question, and it hands back a number you can trust statistically, not a paragraph you have to parse.
bison-1 is the bigger of the pair at 31B dense parameters. groundhog-1 runs as a 26B mixture-of-experts model with roughly 4B parameters active per token, built to be the cheaper option. Both jumped their context window from 16k tokens to 128k, and both now accept images, which matters for anyone running routing, triage, or guardrail pipelines that need to feed in full policy documents or long conversation histories in a single call.
They join Baseten's existing lineup of mammoth-1, elephant-1, and joey-1, and on Artificial Analysis's Decision Index v2.1, bison-1 claims the top spot among 19 ranked models.
On AA Decision Index v2.1, bison-1 scores 76. Liquid AI’s d1.2-ultra scores 75 and mammoth-1 scores 71. The index blends 6 evaluations, led by AA-Paperclip at 30%.
Why this matters
For teams building routing, triage, or guardrail layers, bison-1's weak spot matters as much as its wins. Topping 4 of 6 components on the Artificial Analysis Decision Index is a real result, and a #1 overall score of 76 against 19 models isn't nothing. But that #11 ranking on AA-Omniscience-Decision is the number to sit with before you swap out a production classifier.
Decision models get judged on calibration, not fluency, so a gap like that could mean the model is confidently wrong on exactly the edge cases where guardrails earn their keep. The jump from 16k to 128k context and native image input is the more interesting shift for builders: full policy documents and conversation history in a single call changes what's practical for moderation pipelines, not just what's possible on a leaderboard. Open weights on a Gemma 4 base also means you can actually inspect and fine-tune this thing instead of trusting a vendor's dashboard.
Worth testing against your own failure cases before trusting the aggregate score.
Common Questions Answered
What are decision models and how do bison-1 and groundhog-1 differ from traditional language models?
Decision models like bison-1 and groundhog-1 are built to output calibrated probabilities and numerical scores rather than text, making them ideal for yes/no questions, multiple choice questions, and scoring tasks. Unlike traditional language models that generate prose paragraphs, these decision models provide statistically trustworthy numbers that teams can rely on for routing, triage, and guardrail applications.
What are the parameter sizes and architecture of Baseten's bison-1 and groundhog-1 models?
Bison-1 is the larger model with 31 billion dense parameters, while groundhog-1 operates as a 26 billion parameter mixture-of-experts model. Both models are created through post-training Google's Gemma 4 to specialize in decision-making tasks rather than general text generation.
How does bison-1 rank on the Artificial Analysis Decision Index v2.1 compared to other decision models?
Bison-1 scores 76 on the AA Decision Index v2.1, ranking it above Liquid AI's d1.2-ultra which scores 75 and mammoth-1 which scores 71. The index blends 6 evaluations with AA-Paperclip weighted at 30%, and bison-1 tops 4 of the 6 components, though it ranks #11 on the AA-Omniscience-Decision metric.
Why is calibration more important than fluency when evaluating decision models like bison-1?
Decision models are specifically judged on calibration rather than fluency because their primary purpose is to provide reliable probability estimates and scores for automated decision-making systems. A gap in calibration performance, such as bison-1's #11 ranking on AA-Omniscience-Decision, could indicate the model is confidently producing incorrect predictions, which is critical for production classifiers to identify before deployment.
What use cases are bison-1 and groundhog-1 designed to support?
Bison-1 and groundhog-1 are designed for teams building routing, triage, and guardrail layers that require reliable probabilistic outputs rather than text generation. These decision models excel at automating classification and decision-making tasks where calibrated confidence scores are more valuable than natural language responses.
Further Reading
- Baseten Releases Bison-1 and Groundhog-1, Topping Decision Model Index - MarkTechPost
- Meet bison-1 and groundhog-1, frontier American decision models - Baseten
- bison-1 and groundhog-1 decision models available on Baseten - Baseten
- Bison 1 Benchmarks, Cost and Capabilities - Vals AI
- Bison-1 and Groundhog-1: frontier American-backbone decision models - Base Labs