vLLM Semantic Router Releases Decision 3.0 Open Multimodal Models

vLLM Semantic Router has released Decision 3.0, a new family of open multimodal decision models designed to process text, JSON, and images to answer typed questions. These models are available in five sizes, ranging from 0.8 billion to 27 billion parameters, and are all licensed under Apache-2.0. The primary function of Decision 3.0 is to classify, route, and gate requests by returning probabilities for answers, which developers can then use to set thresholds, offering an alternative to parsing generated text from large language models. This approach is described as the 'system one' pattern for routers and guardrails.

The five models in the Decision 3.0 family are d3-lite (0.85B), d3-nano (2.21B), d3-mini (4.54B), d3-flash (8.39B), and d3 (26.09B). The context length for these models has not been disclosed. Performance metrics indicate that the largest model, the 27B d3, leads both text and vision benchmarks on internal evaluations. Specifically, it achieved a score of 97.7 on KIE (CORD+FUNSD) document extraction, which is considered its best performance. Conversely, its worst performance was recorded at 34.1 on R-Bench-M, also for the d3 model. It is noted that these scores are self-reported against live board data.

Each model in the Decision 3.0 series is a fine-tune of a Qwen base model. The 27B d3 model is built upon Qwen3.8-27B. The smaller models, d3-lite, d3-nano, d3-mini, and d3-flash, are based on Qwen3.5 checkpoints at sizes of 0.8B, 2B, 4B, and 9B, respectively. Every model incorporates a vision encoder. The d3-lite model has a 0.10B vision encoder, while d3-nano and d3-mini feature a 0.33B encoder. The d3-flash and the top-tier d3 models are equipped with a 0.46B vision encoder. These models can process multiple image formats, including PNG, JPEG, and WebP, with each image being read at up to 1.6 megapixels. All provided images are considered for every question asked.

Installation of Decision 3.0 requires transformers version 5.17.0. The optional flash-linear-attention package can be installed to accelerate linear attention computations. The models are designed to run on AMD Instinct MI325X GPUs, with latency measured on this hardware. BF16 weights are used, and no officially quantized variants have been listed at the time of release. The development team highlights that a 9B model within this generation now surpasses the performance of last generation's 27B models, indicating significant efficiency gains. The system allows developers to pass a state, which can include text or JSON, along with optional images and named questions. Each question can be a choice between options, a Yes/No query, or a score on a defined scale. The model answers all questions in a single call and provides a probability for each answer, facilitating programmatic decision-making.