bison-1
AvailableDecision model from Base Labs at Baseten, released Nov 20 2026: a full fine-tune of Gemma 4 31B that takes a state (text, optionally images) plus a set of questions and returns a calibrated probability for every answer option from one shared-prefix prefill, without generating text. Block-bidirectional attention on full-attention layers, 131,072-token max sequence length. Baseten reports #1 of 19 decision models on the Artificial Analysis Decision Index v2.1 (76, ahead of Liquid d1.2-ultra at 75) and the best calibration on that index (Brier 0.14). Available on Baseten Model APIs as decision-models/baseten-bison-latest at $0.10 per 1M tokens, with open weights on Hugging Face (MIT and Apache-2.0).
Specifications
- License
- Open source · MIT and Apache-2.0
- Weights
- Downloadable
- Architecture
- Dense
- Parameters
- 31B
- Context window
- 131K tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- $0.1 / ?
- Modalities
- TextVision
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.