vLLM Semantic Router Team Releases Decision 3.0: Open Multimodal Models Ranked #1 on Jev Decision Index

The vLLM Semantic Router team has released Decision 3.0 (d3), a new family of open-source, multimodal decision models ranging from 0.6B to 27B parameters. Licensed under Apache-2.0, these models are designed for efficient decision-making tasks rather than text generation, supporting text, images, and video inputs.

Architecture and Capabilities

Built on Qwen3.5 (0.8B, 2B, 4B, 9B) and Qwen3.8-27B backbones, the models include vision encoders within their parameter counts (e.g., 0.33B in the 4B model, 0.46B in the 27B model). They support three decision types: choice, yes/no, and score, providing probabilities for every answer without generating text. Multiple questions about the same input can be processed in a single call via separate forward passes.

• Images: Supports multiple images per request, read at up to 1.6 MP.
• Video: Supports multiple videos per request (MP4, WebM, MOV, MKV), sampled at 2 fps with a maximum of 32 frames per clip, each at up to 0.2 MP.
• Weights: BF16 format, loadable via Transformers with trust_remote_code=True.

Benchmark Performance

In evaluations using the official Jev 0.3 kit, Decision 3.0 achieved #1 rankings on both the text and vision boards of the Jev Decision Index:

• The 27B model scored 64.7 on text and 71.6 on vision, surpassing all listed models, including Jev's own.
• Every model size leads its respective class: the 0.6B, 0.8B, 4B, and 9B models each top their size category.
• The 9B model (d3-flash) outperforms the previous generation's 27B model (Decision 2.0).

Video Zero-Shot Results

Despite no video data being used during training, the models demonstrate strong zero-shot video understanding on the Perception Test validation set (19,140 questions):

• The 27B model scored 77.4.
• The 4B model scored 70.3.

The models are available in the Hugging Face collection: vllm-sr/decision-30.