vLLM Semantic Router Releases d3-ultra Decision Model

This decision model is designed to read an input once and return a probability for every answer to every question, enabling developers to perform routing, classification, visual grounding, and video checks through a single open-weight call.

d3-ultra has 26.09 billion parameters, including a 0.46 billion vision encoder, and supports up to 131,072 tokens of context per question.

The model was tested on one AMD Instinct MI325X with BF16 weights, and no official quantized variant has been listed yet.

In private testing by the team, d3-ultra achieved the highest SR-Text score among six decision models, with a best score of 97.8 on SR-Text tools and automation.

Its performance on SR-Box was 54.6% across nine domains, which was 5.0 points lower than its sibling model, d3-turbo.

d3-ultra is part of the Decision 3.0 family, which ranges from 0.6B to 27B parameters, and it answers typed questions with probabilities.

Users can send a state, which can include text, JSON, images, or videos, along with a set of questions that can be of type Choice, Yes/No, Score, Box, or Track.

The model's core principle is 'state once,' meaning it reads the input one time per request to answer all questions.

The architecture features a Qwen3.5-architecture backbone with 64 layers, comprising 48 linear-attention layers and 16 full-attention layers.

Box answers consistently return one box with a confidence score, and questions exceeding 131,072 tokens will result in a max_length_exceeded error.