Is bison-1 state of the art?Pending benchmarks.Announced Nov 20, 2026; no independent results yet.
Base Labs' 31B decision model, post-trained from Gemma 4 31B IT. Official model card: it takes text and optional images and a set of questions, and returns calibrated option probabilities in a single shared-prefix prefill without generating text, with a 131,072-token context window. Ungated open weights are on Hugging Face; it is live on Baseten Model APIs as decision-models/baseten-bison-latest. Official price is $0.10 per 1M tokens.
The details so far
- Lab
- baselabs
- Announced
- Nov 20, 2026
- API
- Live · decision-models/baseten-bison-latest
- Open weights
- Published · Ungated open weights on Hugging Face at baselabs/bison-1 (Base Labs model card)
- Parameters
- 31B
- Type
- Official model card: a decision model post-trained from Gemma-4-31B by Base Labs at Baseten. It answers typed questions (yes/no, choice, and ordered score) from one shared-prefix prefill and returns calibrated probabilities without generating free-form text. Inputs are text and images. Attention uses a block-bidirectional mask on full-attention layers; sliding-window layers stay causal (Base Labs model card and Baseten library)
- Context
- Official model card and Baseten library: 131,072 tokens (Base Labs model card and Baseten library)
- Hosted access
- Official model card and Baseten library: live on Baseten Model APIs as decision-models/baseten-bison-latest, using the same API as mammoth-1 and elephant-1, at https://inference.baseten.co/v1/systemone (Base Labs model card and Baseten library)
- Pricing
- Official model card and Baseten library: $0.10 per 1M tokens (Base Labs model card and Baseten library)
- License
- Baseten's modifications are released under the MIT License; the Gemma 4 31B IT base model is under Apache 2.0. Use of the model is subject to both licenses (Base Labs model card and LICENSE)
- Related
- Dense sibling of groundhog-1; separate from mammoth-1, elephant-1, and joey-1
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: self-reported score of 76 on the Artificial Analysis Decision Index v2.1, scored by Artificial Analysis on November 19, 2026; the lab lists that score ahead of d1.2-ultra (75) and mammoth-1 (71) among 19 decision models on that index, with a Brier score of 0.14 (official bison-1 model card).
Lab claim: self-reported 84.8 on LongMemEval M and 87.1 on LongMemEval S at 128k tokens in the lab's own run (official bison-1 model card).
Sources
LABbison-1 (Base Labs model card)huggingface.co/baselabs/bison-1LABbison-1 (Baseten Model APIs library)baseten.co/library/bison-1/Compare when it lands
Unscored: groundhog-1. A scored comparison appears once both are on a shared board.
Tracking continues until an independent board lists it.