What's Happening?
Baseten, an AI inference platform, has released two new open-weight decision models, bison-1 and groundhog-1, built on Google's Gemma 4 architecture. These models are designed to provide calibrated probabilities for yes/no, choice, and score questions
rather than generating text. The bison-1 model has achieved a score of 76 on the Artificial Analysis Decision Index v2.1, placing it at the top among 19 models. Both new models significantly expand context windows from 16k to 128k tokens and are capable of accepting image inputs. Bison-1 is a 31B dense model, while groundhog-1 is a 26B Mixture-of-Experts (MoE) model with approximately 4B active parameters. They are available on Baseten Model APIs, with bison-1 priced at $0.10 per 1 million tokens and groundhog-1 at $0.035 per 1 million tokens. The company attributes bison-1's superior calibration to its Gemma base and the use of 'soft labels' in its training mix, resulting in a Brier score of 0.14, which is the lowest and best on the index.
Why It's Important?
The introduction of Baseten's bison-1 and groundhog-1 models marks a significant advancement in the field of AI decision-making, particularly for applications requiring high accuracy and calibrated probabilities. By topping the Artificial Analysis Decision Index, bison-1 demonstrates a leading capability in tasks such as ticket triage, spam detection, jailbreak guardrails, and LLM routing. The expanded context window of 128k tokens allows for more comprehensive policies and histories to be processed in a single call, enhancing the efficiency and effectiveness of AI pipelines. The competitive pricing structure, especially for groundhog-1, makes advanced decision-making AI more accessible for developers and businesses. This development could lead to more reliable and nuanced AI-driven automation across various industries, reducing errors and improving operational efficiency in critical decision-making processes.
What's Next?
Developers can immediately deploy bison-1 and groundhog-1 via Baseten Model APIs, with a straightforward transition from older decision models by simply changing the model name. The open BF16 weights for both models are also available on Hugging Face, allowing for broader experimentation and integration within the AI community. Baseten anticipates that groundhog-1 will eventually match the performance of its mammoth-1 model on official indices. The company will likely continue to refine these models and potentially introduce new features or optimizations based on user feedback and performance data. The success of these models could also spur further innovation in decision-focused AI, prompting competitors to develop more accurate and cost-effective solutions, thereby accelerating the overall progress in AI deployment and application.
Beyond the Headlines
The emphasis on 'calibrated probabilities' and 'soft labels' in the training of bison-1 highlights a crucial shift in AI development towards models that not only make decisions but also accurately convey their uncertainty. This capability is vital for building trust in AI systems, especially in sensitive applications where understanding the confidence level of an AI's output is as important as the output itself. The ability to process images alongside text in decision models opens up new possibilities for multimodal AI applications, allowing for more holistic and context-aware decision-making. Furthermore, the release of open-weight models fosters transparency and collaboration within the AI research community, potentially accelerating the development of even more advanced and robust AI solutions. This move by Baseten contributes to the broader trend of democratizing advanced AI capabilities, making them accessible to a wider range of developers and organizations.













