New Multimodal Decision Models Released by vLLM Semantic Router Team

The vLLM Semantic Router team has announced the release of Decision 3.0 (d3), a new family of multimodal decision models. The release includes models ranging in size from 0.6 billion to 27 billion parameters, designed to process text, images, and video. According to documentation provided by the developers, the models are released under the Apache-2.0 license and are available via Hugging Face.

Technical Architecture and Capabilities

The Decision 3.0 family is built upon the Qwen3.5 architecture for smaller models (0.8B, 2B, 4B, and 9B) and the Qwen3.8-27B architecture for the largest iteration. The parameter counts provided by the developers include the integrated vision encoder, which accounts for 0.33B in the 4B model and 0.46B in the 27B model.

Unlike standard large language models that focus on text generation, the d3 models are specialized for decision-making tasks, specifically providing “choice,” “yes/no,” or “score” outputs. The developers note that the models provide a probability for every answer without generating accompanying text. The architecture allows for multiple questions regarding a single input to be answered in one call through individual forward passes.

The models support multimodal inputs, including:

  • Images: Multiple images per request, with a resolution of up to 1.6 MP.
  • Video: Support for MP4, WebM, MOV, and MKV formats, sampled at 2 frames per second with a maximum of 32 frames per clip at 0.2 MP each.

The models were trained on AMD Instinct MI325X hardware and are available in BF16 weights.

Performance Benchmarks

According to internal evaluations conducted by the developers using the Jev 0.3 kit, the 27B model achieved a score of 64.7 in text and 71.6 in vision, ranking first on both Jev Decision Index boards. The developers claim that every model size in the series leads its respective class. Notably, the 9B model, referred to as “d3-flash,” reportedly outperformed the previous-generation Decision 2.0 27B model.

In zero-shot video testing—where the developers state no video was included in the training data—the 27B model scored 77.4 and the 4B model scored 70.3 on the 19,140-question Perception Test validation set.

Market and Implementation Context

While the technical release marks a significant update to the Decision series, discussions on community platforms like Reddit reflect a divide between technical utility and enterprise adoption. Some users noted that while the models offer immediate utility for specific tasks, widespread adoption in corporate environments—particularly in legal departments—may be delayed by months due to internal approval processes.

The release also highlights the ongoing hardware demands associated with running large-scale models. Users in the developer community continue to evaluate complex hardware configurations, such as multi-GPU setups involving RTX 3090 and 4090 cards, to support concurrent agent workflows and long-context tasks required by modern LLM architectures.