Decision 3.0: Open Multimodal Decision Models Ranging from 0.6B to 27B
The vLLM Semantic Router team has released Decision 3.0, a new family of open-source multimodal decision models ranging from 0.6B to 27B parameters under the Apache-2.0 license. These models process text, images, and video, achieving the number one ranking on both Jev Decision Index leaderboards. This release provides state-of-the-art performance across various parameter sizes, allowing efficient deployment of decision-making agents that output structured choices instead of free-form text. It pushes the boundaries of open-source multimodal intelligence, offering superior zero-shot video capabilities without prior video training. Decision 3.0 models are built on Qwen architectures and feature three specialized decision types—choice, yes/no, and score—outputting precise probabilities without text generation. The 27B model scored 71.6 on vision and 64.7 on text, while demonstrating zero-shot video capabilities across all 19,140 Perception Test validation questions.
## BACKGROUND
The Jev Decision Index is a benchmark suite used to evaluate multimodal decision models using a combination of public benchmarks, private skill tests, and private tasks from new domains. Open-source decision models focus on classification and structured choices rather than conversational text generation, making them highly efficient for routing and decision-making pipelines.