Baseten Launches Two Open-Weight Decision Models Based on Google’s Gemma 4, Topping Benchmarks in Their Category
Large language model deployment service provider Baseten has released two decision models: bison-1 and groundhog-1, both built on Google’s Gemma 4, supporting 128k context windows and image input. Notably, bison-1 scored 76 in the AA Decision Index v2.1 (a benchmark for professional decision-making capabilities), ranking first among models in its class with open weights enabling private deployment.
Large language model deployment service provider Baseten has recently released two open-weight decision models, bison-1 and groundhog-1, whose base models are powered by Google’s latest open-source large language model, Gemma 4. Among them, bison-1 scored 76 in the AA Decision Index v2.1 test, topping the list of decision models in its segment. The AA Decision Index is an industry benchmark test set specifically for evaluating the decision-making capabilities of large language models, covering real business scenarios such as multi-condition logical judgment, risk-reward trade-offs, and business plan selection. The score directly corresponds to the reliability of the decisions output by the model. Currently, mainstream general-purpose large language models mostly score between 50 and 65 in this test. Both models support 128k context windows and image input. A 128k context can accommodate approximately 90,000 words of text at once, allowing entire 100-page business contracts or sets of quarterly business ledgers to be fully input without splitting into segments. The image input capability supports the model to directly recognize non-text information such as screenshot documents, data charts, and on-site photos. For example, uploading a photo of an inventory count sheet can directly generate corresponding replenishment scheduling suggestions. These decision models are specially optimized for business judgment scenarios, cutting off the redundant chat and creative generation capabilities of general-purpose large language models, resulting in lower inference costs and outputs more aligned with standardized business rules. The open-weight design allows enterprises to download model weights for private deployment, so core business data does not need to leave the internal network, adapting to scenarios with high data privacy requirements, such as financial risk control, industrial production scheduling, and supply chain management. Currently, the official only discloses basic parameters and benchmark scores, and has not yet revealed specific scenario differences between the two models. The information comes from a report by tech media MarkTechPost.
发布时间: 2026-11-21 12:01