Baseten Releases bison-1 Decision Model
Baseten's open-weight bison-1 and groundhog-1 models target decision-making tasks, with bison-1 leading a 19-model index.
Updated on Nov. 21, 2026 in Artificial Intelligence

Baseten has released bison-1 and groundhog-1, two open-weight decision models built on Google's Gemma 4. Both models extend context length from 16k to 128k tokens and accept images as input.
Why it matters
The release gives developers open-weight models aimed at decision tasks, where calibrated probability estimates matter alongside raw accuracy. bison-1's ranking on an independent index offers a measurable point of comparison for teams choosing among models.
bison-1 is a 31B dense model and groundhog-1 is a 26B mixture-of-experts model with about 4B active parameters per token, both extended to 128k-token context. bison-1 scores 98 on When2Call and 60 on AA-Omniscience-Decision v1.1, ranking #11, while groundhog-1 has no official index score yet.
The players
Baseten
Baseten is an AI inference infrastructure company that hosts and serves machine learning models for developers.
Google is the technology company that developed the Gemma model family on which bison-1 is built.
Artificial Analysis
Artificial Analysis is an independent benchmarking organization that publishes the Decision Index used to rank these models.
The details
Both models answer each question in a single prefill pass and return P(yes) or the chosen option with a probabilities field. Requests longer than the trained limit are refused rather than truncated. Baseten credits the Gemma base and additional soft-label training data for bison-1's calibration.
Timeline
Baseten released bison-1 and groundhog-1 on November 20, 2026.
The Tech Race
The release places bison-1 at the top of the Artificial Analysis Decision Index v2.1, which ranks 19 models on decision-making performance. It enters a competitive field where open-weight releases from smaller labs are now benchmarked directly against larger proprietary systems.
Developers can call bison-1 at $0.10 per 1M input tokens and groundhog-1 at $0.035 per 1M input tokens, lowering the cost of running decision workloads. Teams that need calibrated probabilities can test bison-1 directly against their own tasks before switching.
The takeaway
Calibration and context length are becoming the points on which open-weight decision models compete, so teams should test a model's probability outputs on their own data before adopting it. Benchmark rankings are a starting point rather than a substitute for task-specific evaluation.
Further reading
Explore more Artificial Intelligence coverage on the models shaping automated decision-making.
Source note: This article includes information reported by MarkTechPost.



