Baseten says its open-weight bison-1 leads decision model index with 76 points
Baseten has released two decision models based on Gemma 4 that do not write text but give a probability for every answer option. The company says bison-1 scored 76 on the Artificial Analysis index, but the cheaper groundhog-1 has not yet been scored.
Baseten released two decision models on 20 November. bison-1 has 31 billion parameters. groundhog-1 is a 26-billion-parameter mixture-of-experts (MoE) model in which about 4 billion parameters process each token. Both are post-trained from Google’s Gemma 4 models, and their model weights can be downloaded from Hugging Face.
A decision model receives a situation, meaning text and optionally images, and a set of questions. A question can be yes/no, a choice between named options or a level on an ordered scale. The model answers all the questions in a single pass over the input and returns a probability for every option. In an example from the groundhog-1 model card, a customer’s message about a double charge prompts three questions at once: whether the customer wants a refund, which team the message belongs to and how urgent it is.
A calibrated model’s probabilities can be trusted: when the model gives an answer a 70 per cent probability, it should be right about seven times out of ten. The Brier score measures the gap between predicted probabilities and actual answers, and a lower score is better. MarkTechPost summarised Baseten’s release post and the bison-1 model card. It reported that Baseten gives bison-1 a Brier score of 0.14 and puts Liquid AI’s d1.2-ultra at 0.15. The company attributes bison-1’s good calibration to the Gemma base and to training data with more examples in which the target is a probability rather than a single correct answer.
The table shows Artificial Analysis Decision Index v2.1 scores, context windows and prices from MarkTechPost’s comparison.
| Model | Index v2.1 | Context window (tokens) | Price / 1M input tokens |
|---|---|---|---|
| bison-1 | 76 | 131,072 | $0.10 |
| groundhog-1 | not scored | 131,072 | $0.035 |
| mammoth-1 (Baseten) | 71 | 16k | $0.10 |
| Liquid AI d1.2-ultra | 75 | 65,536 | $0.042 |
Source: MarkTechPost. Index scores are Baseten’s and Liquid AI’s own figures.
The gap to d1.2-ultra is one point, and MarkTechPost notes that all the figures are self-reported by the companies. Baseten expects groundhog-1 to reach the level of mammoth-1, but the index has not yet scored it. Of the index’s six component tests, bison-1 ranks lowest on the knowledge-heavy AA-Omniscience-Decision test, scoring 60 and placing 11th. In Baseten’s reported LongMemEval M result with 128,000-token inputs, bison-1 scored 84.8 and groundhog-1 77.4.
Requests longer than the trained length limit are rejected, not truncated. Systems that assume input will be silently truncated therefore need a length check, MarkTechPost said.
The models run on Baseten Model APIs with the same API as the company’s earlier decision models, so switching only requires changing the model name to decision-models/baseten-bison-latest or decision-models/baseten-groundhog-latest. MarkTechPost calculates that the BF16 weights take up about 62 and 52 gigabytes respectively. Baseten’s changes are released under the MIT licence and the base model under the Apache 2.0 licence. MarkTechPost’s service listing, however, gives the context for Baseten’s API as 33K, which does not match the stated 128k.
Read more: Models