Baseten has released two new open-weight models designed to make structured choices rather than write conventional chat answers. Announced on November 20, bison-1 and groundhog-1 join the Base Labs decision-model family. Both accept text and images and are offered through Baseten Model APIs, while their official Hugging Face repositories expose the weights for teams interested in inspecting or hosting the checkpoints. Primary source ↗
The two models make different hardware and performance trade-offs. Baseten describes bison-1 as a dense 31-billion-parameter model built on Google Gemma 4 31B. groundhog-1 uses the Gemma 4 26B-A4B mixture-of-experts architecture, with approximately four billion active parameters per token. Both offer a 128K context window through the announced APIs. Active parameter counts should not be mistaken for total memory required to store an MoE checkpoint.
Decision models evaluate explicitly defined answer choices and assign probability estimates. This makes them attractive for routing, prioritization, quality checks and similar applications where a system must select one of several actions. For example, an operations workflow could present a model with policy documents and ask which review queue a case belongs in. Long context may reduce the need to discard relevant history, but it does not by itself establish better decisions.
Baseten says bison-1 scored 76 on Artificial Analysis Decision Index v2.1, ahead of 18 other evaluated decision models at the time of its announcement. The company also reports a Brier score of 0.14, a calibration metric where lower values indicate closer correspondence between predicted confidence and observed outcomes. These numbers describe a particular published benchmark and evaluation setup. Open Model Weights has not reproduced them, and groundhog-1 does not yet have an official index score in Baseten's announcement.
Licensing deserves the same scrutiny as performance. The two official checkpoint pages identify their terms as a combination of MIT and Apache 2.0 rather than a single blanket permission. Anyone redistributing or embedding the models should examine the repository license documents and inherited base-model terms. Public weights are useful evidence of availability, but do not resolve every downstream packaging, dependency or deployment question.
The release illustrates a meaningful specialization within open-weight AI: probabilistic decisions can be evaluated for calibration, error cost and latency rather than only general chat quality. For serious adoption, teams should test confidence reliability on their own distribution, compare decisions against a documented policy baseline and measure runtime costs under a pinned checkpoint revision. The headline benchmark is a lead for investigation, not a substitute for that work.
What this changes for the evidence layer
Official Base Labs checkpoint pages provide downloadable weights and list a combined MIT-and-Apache-2.0 license designation. Before canonical model intake, verify immutable model revisions, complete artifacts, exact layered license text and runtime requirements. Artificial Analysis Decision Index and Brier figures are publisher-attributed; OMW has not run independent evaluation.
Baseten Base Labs — official bison-1 and groundhog-1 release announcement
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗