Baseten has released bison-1 and groundhog-1, two open-weight decision models post-trained on Google's Gemma 4 that output calibrated probabilities for yes/no, choice, and score questions without generating text tokens. Bison-1, a 31B dense model, tops the Artificial Analysis Decision Index v2.1 with a score of 76, edging out Liquid AI's d1.2-ultra (75) and Baseten's prior Qwen-based mammoth-1 (71). It achieves the index's lowest Brier score of 0.14, attributed to Gemma 4's base quality and a training mix incorporating soft probability labels. Component scores show bison-1 ranking first on four of six evaluations, including When2Call (98), EvalSafe-O*NET-AA (93), AA-Symmetry v1.1 (89), and HLE-MultiChoice v1.1 (51), while scoring 73 on AA-Paperclip (rank #2) and 60 on AA-Omniscience-Decision v1.1 (rank #11). Both models extend context from 16k to 128k tokens (131,072) and accept image inputs, enabling full policy and history ingestion for routing, triage, and guardrail pipelines. Groundhog-1, a 26B MoE model with ~4B active parameters, targets cost efficiency at $0.035 per 1M input tokens versus bison-1's $0.10 (matching mammoth-1). On LongMemEval M (128k), bison-1 scores 84.8 and groundhog-1 77.4. Both models are available via Baseten Model APIs under MIT + Apache 2.0 licenses with open BF16 weights on Hugging Face (approx. 62 GB and 52 GB respectively); no quantized builds or inference GPU specs are listed. Requests exceeding the context limit are refused rather than truncated.
Read original
marktechpost