Skip to content
Neural Daily

ModelsNews

d3-ultra: vLLM's open 27B multimodal decision model

🇭🇺 Olvasd el magyarul →

vLLM Semantic Router released d3-ultra, a 27-billion-parameter multimodal decision model under Apache-2.0. It does not generate text, but returns a probability for every possible answer.

By: Mikola Máté Patrik · · 2 min read

d3-ultra banner reading Towards Open Multimodal Foundation Decision Models over a landscape painting
Image: vLLM Semantic Router · Press kit · Official model card banner, background painting by Frederic Edwin Church, CC0 · Source

What is it?

The vLLM Semantic Router project released d3-ultra, the largest member of its Decision 3.0 family (Hugging Face). The model has 26.09B parameters (including a 0.46B vision encoder) and an Apache-2.0 license. It accepts text, JSON, images and videos, with 131,072 tokens of context per question.

The key point is that it does not generate text. It answers by question type: choice between options, yes or no, a score on a scale, a bounding box in an image or video moment, or a track across a whole video. Every answer comes with a probability, so uncertainty can be read directly.

Why is it fast?

The model reads the input once and answers every question from that single read. On one AMD Instinct MI325X, one question takes 304 ms and 255 questions about the same photo take 2.16 seconds (median).

How good is it?

On the project's own SR-Text benchmark it scores 87.2, ahead of GLiDE V1.2 at 85.1, the earlier d1-ultra at 83.4 and Jev 1.14 at 73.1. In visual localization (SR-Box, 10 domains) it reaches 58.2 percent, below the smaller d3-turbo at 60.0 percent but well above Jev 1.14 at 45.7 percent (MarkTechPost).

d3-ultra SR-Text results table by capability compared with other decision models
vLLM Semantic Router · Press kit · Official evaluation table from the model card · Source

Source: vLLM Semantic Router. SR-Text results by capability, self-reported.

What is it good for?

Routing, moderation, ticket classification or quality checks, where you need a reliable decision and a probability rather than text. It runs with vllm-sr serve or Transformers 5.17.0, and there is a free API at decision.vllm-sr.ai. The benchmarks are self-reported, and SR-Text is the project's internal test.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles