The vLLM Semantic Router team has released d3-ultra, a 27B open-weight multimodal decision model that works very differently from a chatbot: it does not generate text. Instead, it reads an input once and returns a probability for every answer to every question — enabling routing, classification, visual grounding and video checks from a single open-weight call.
d3-ultra (Credits: MarkTechPost)
At 26.09B dense parameters, including a 0.46B vision encoder, d3-ultra is the newest member of the Decision 3.0 family, which spans 0.6B to 27B. It is released under Apache-2.0, with 131,072 tokens of context per question, and has been tested on a single AMD Instinct MI325X GPU with BF16 weights. No official quantized variant is listed yet.
d3-ultra: One Read, Every Answer
The core idea is “state once”: you send a state — text or JSON, plus images or videos — and a set of questions. Each question can ask the model to locate something in an image or a video moment, choose among options, or classify the input. The model reads the input once per request and answers every question from that single read.
That design makes it suited to practical pipelines: ticket routing, SQL checks, tool and automation decisions, and visual verification, where a system needs a confident decision rather than a paragraph of generated text. Box answers always return exactly one box with a confidence score — a confidence near zero means the target is not in the picture.
d3-ultra Benchmarks: Strong on Text, Mixed on Visual Grounding
On the team’s own SR-Text test — 10,411 text requests with 73,495 questions, scored with chance-corrected accuracy — d3-ultra scores 87.2, the highest of the six decision models compared, ahead of d3-turbo (83.1) and GLiDE V1.2 (85.1). It leads every area except arts and human taste, where its sibling d1-ultra edges it 90.8 to 89.6, within the margin of error.
Visual grounding is more mixed. On SR-Box, with 5,678 box questions across 4,984 images, d3-ultra scores 58.2% over 10 domains and 54.6% over 9 domains — about five points below d3-turbo. The team says d3-turbo stays stronger on natural images, while d3-ultra leads on charts.
Latency is a selling point: on one MI325X, one question about a photo takes 304 ms, while 255 questions take 2.16 seconds — about 7.3 ms per extra question. The model is available via Transformers and the vLLM Semantic Router engine, with a free API listed at decision.vllm-sr.ai. For developers building decision-heavy AI systems, d3-ultra is less a chatbot replacement than a fast, open-weight judge: one read, probabilities for everything.
Source: MarkTechPost — read the full story there.
