vLLM Open-Sources d3-ultra: A 27B Probabilistic Decision Model for Multimodal Single-Pass Inference
The vLLM Semantic Router team has open-sourced d3-ultra, a 26-billion-parameter probabilistic decision model. Utilizing a 'state once' architecture, it processes input in a single pass to simultaneously handle diverse tasks like multiple-choice questions, bounding box detection, and video tracking, achieving high efficiency on AMD MI325X hardware. While it outperforms competitors on text semantic routing benchmarks, it trails its lighter sibling in box grounding accuracy, with some test sets remaining private.
Full Report
▲ Marktechpost AI · 2026-10-27 05
▲ vLLM Semantic Router just open-sourced d3-ultra, a 27B decision model that answers questions with probabilities instead of text.
The key idea is "state once." The model reads your input a single time, then answers every question on it. Those questions can be choices, yes/no, scores, boxes in an image, or a track across a video.
On the team's own SR-Text test, d3-ultra scores 87.2, ahead of Fastino's GLiDE V1.2 at 85.1 and Liquid AI's d1-ultra at 83.4. On 1 AMD Instinct MI325X, 255 questions about 1 photo take 2.16 s. It is a trade-off though: on Box grounding it trails its sibling d3-turbo, 58.2% vs 60.0%, and both test sets are private.
Weights are on Hugging Face under Apache-2.0: 26.09B parameters, 131,072 tokens per question, text, JSON, images and video in......
Full analysis:
