Launch of open decision models
Baseten has launched Bison-1 and Groundhog-1, two open-weight models designed specifically for decision tasks. These models were post-trained from the Gemma 4 family by Google, marking a shift from the company's previous models, which were based on the Qwen architecture.
The release took place on November 20, 2026, and is immediately available through Baseten Model APIs. Additionally, the weights for both models have been released on Hugging Face under an open license, allowing teams to audit the training and deploy inference on their own infrastructure if desired.
Architecture and Gemma 4 base
Bison-1 is built on Gemma 4 31B, a dense model with 31 billion parameters. Groundhog-1, on the other hand, is based on Gemma 4 26B-A4B, a mixture-of-experts (MoE) architecture with 26 billion total parameters, of which only 4 billion are active per token. This architectural difference allows Groundhog-1 to offer a cost-to-latency balance that approaches that of smaller models while maintaining superior reasoning capability.
The use of Gemma 4 responds to the demand from teams seeking models with a Western base for technological sovereignty and regulatory compliance. The Gemma 4 family, launched by Google, is available under the Apache 2.0 license, which facilitates its integration into corporate environments that require traceability of the base model's origin.
Technical comparison of Baseten's decision model family
| Model | Parameters | Base Architecture | Context Window | Price per 1M tokens |
|---|---|---|---|---|
| Bison-1 | 31B | Gemma 4 31B | 128k | $0.10 |
| Groundhog-1 | 26B (MoE, ~4B active) | Gemma 4 26B-A4B | 128k | $0.035 |
| Mammoth-1 | 27B | Qwen | 16k | $0.10 |
| Elephant-1 | 9B | Qwen | 16k | $0.035 |
Context and multimodal capabilities
One of the most significant improvements is the expansion of the context window to 128k tokens, which represents an eight-fold increase compared to the 16k tokens of Baseten's previous models. This capability allows including complete documents, extensive email threads, or full corporate policies in the model input without the need to summarize or select fragments.
Both models accept multimodal input, processing text and images simultaneously. This is particularly useful in scenarios where the decision depends on scanned documents, damage photographs, or user interface screenshots. The combination of long context and vision allows the model to capture details that may be far from the main question or that are only visible in the graphical component.
Performance and calibration on the Decision Index
On the Decision Index v2.1 by Artificial Analysis, Bison-1 achieves a score of 76, placing it first among 19 evaluated decision models. It is followed by d1.2-ultra with 75 points and Mammoth-1, Baseten's previous model, with 71 points. Baseten expects Groundhog-1 to reach comparable performance to Mammoth-1 on this index.
A highlighted aspect is the probability calibration. Bison-1 presents the lowest Brier score on the index, with a value of 0.14, compared to 0.15 for d1.2-ultra. A lower Brier score indicates that the model's stated confidence aligns better with its empirical accuracy. This was achieved in part through the use of soft labels during training, teaching the model to assign probabilities such as 83% for option A and 17% for B, rather than binary 100% labels.
Availability and deployment considerations
The models are immediately available through the Baseten Model APIs and as open weights on Hugging Face. The API maintains compatibility with previous models, meaning teams already using Mammoth-1 or Elephant-1 can migrate simply by changing the model name in their requests, without needing to rewrite integration logic.
For teams opting for self-hosting, the open weights allow implementation in VPC or local environments. However, this requires managing GPU infrastructure, inference batching, and monitoring. Baseten offers the managed option for those who prefer to avoid operational complexity, while the open weights option provides flexibility for audits and data residency requirement compliance.
Steps to evaluate migration to Bison-1 or Groundhog-1
- 01Establish a baseline with the current decision model (rules, embeddings, or chat model).
- 02Mirror real traffic on Baseten using Bison-1 on a representative sample of cases.
- 03Compare calibration metrics (Brier, ECE) and accuracy on your own labels, not just aggregated indexes.
- 04Perform load tests with 128k-token inputs to verify that p95 latency meets SLAs.
- 05Evaluate the need for multimodality and run ablation tests to measure the actual contribution of vision.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction