vLLM Semantic Router launches d3-ultra, a 27‑B open multimodal decision model

vLLM Semantic Router announced the release of d3-ultra, a 27 billion‑parameter open‑weight multimodal decision model that can ingest text, images and video in a single forward pass and output probability scores for every possible answer. The model achieved an 87.2 % score on the team’s SR‑Text benchmark, positioning it among the most capable open‑source decision‑oriented models currently available, according to the announcement on MarkTechPost【https://www.marktechpost.com/2026/10/26/vllm-semantic-router-d3-ultra-27b-multimodal-decision-model/】.

---

What is the d3-ultra model and why does it matter?

d3-ultra is a multimodal decision model built on the vLLM Semantic Router framework. Unlike typical large language models that generate text, a decision model evaluates a set of candidate answers and returns a probability distribution over them. The key differentiator for d3-ultra is its ability to process three modalities—text, images, and video—simultaneously in a single inference step. This reduces latency and resource consumption compared with pipelines that chain separate encoders for each modality. The 27 B parameter count places it in the same scale as many commercial foundation models, yet the weights are released openly, allowing anyone to run or fine‑tune the model without licensing restrictions.

---

How was d3-ultra evaluated?

The developers of the Semantic Router suite evaluated d3-ultra on an internal benchmark called SR‑Text, which measures a model’s ability to select the correct answer from a list of candidates across a variety of textual reasoning tasks. d3-ultra scored 87.2 %, a figure the team highlighted as evidence that the model’s decision‑making capabilities translate well from pure text to multimodal contexts. The benchmark is not publicly documented, but the reported score suggests competitive performance relative to other open‑source decision models that typically hover in the 70‑80 % range.

---

Who can benefit from a multimodal decision model?

The model’s design makes it attractive to three overlapping audiences:

• Creators who need rapid content classification—e.g., automatically tagging video frames, extracting key objects from images, or selecting the most relevant caption for a piece of text.
• Developers building AI‑augmented applications that must fuse visual and textual cues, such as conversational agents that answer questions about a video clip or image gallery.
• Researchers exploring multimodal reasoning, who can now experiment with a large, openly available model that provides probability‑based outputs rather than generative text, simplifying analysis of model confidence.

Because d3-ultra processes all modalities in one pass, it can be embedded in real‑time pipelines where latency matters, such as interactive chatbots, video‑search engines, or on‑device assistants.

---

How does d3-ultra compare to existing models?

Most publicly released multimodal models fall into two categories:

1. Encoder‑only models (e.g., CLIP) that produce embeddings for each modality but leave the decision logic to downstream classifiers.
2. Generative multimodal models (e.g., Flamingo, Gemini) that generate text conditioned on visual inputs.

d3-ultra bridges the gap by combining multimodal encoding with a decision head that directly outputs answer probabilities. This architecture reduces the engineering overhead of stitching together separate encoders and classifiers. While the 27 B size is larger than many open‑source encoders, the open‑weight release offsets the cost by eliminating subscription fees.

---

How to try d3-ultra with Graf by?

Graf by’s Chat app lets users interact with a range of AI models side‑by‑side. By adding the d3-ultra checkpoint to the model catalogue, creators can experiment with multimodal queries directly in the chat interface—upload an image, paste a video URL, and ask a multiple‑choice question. The model will return a probability distribution for each answer, making it easy to see its confidence levels.

For developers who want to prototype more complex workflows, the Graph tool can be used to build a branching prompt chain: feed a video frame into d3-ultra, route the highest‑scoring answer to a text‑generation model for elaboration, and then visualize the decision path. This visual approach helps teams iterate on prompt engineering without leaving the Graf by environment.

If users simply want to read the original paper or related technical notes, the Reader app can ingest PDFs and allow question‑answering over the document, complementing d3-ultra’s multimodal capabilities.

---

What are the practical limitations and next steps?

While d3-ultra’s open‑weight release is a major step forward, there are a few practical considerations:

• Hardware requirements – Running a 27 B model with video input still demands a GPU with substantial VRAM (≥ 24 GB) or a multi‑GPU setup, which may be out of reach for hobbyists.
• Benchmark transparency – The SR‑Text benchmark is internal; independent validation will be needed to confirm the reported 87.2 % score across diverse tasks.
• Fine‑tuning – The model can be fine‑tuned on domain‑specific multimodal datasets, but the process can be costly. Graf by’s Code studio could help automate the fine‑tuning pipeline with containerised training scripts.

Future releases from the vLLM team may include smaller, distilled versions of d3-ultra for edge devices, as well as broader benchmark reporting to aid adoption.

---

Frequently asked questions

How does a decision model differ from a generative model?

A decision model evaluates a fixed set of candidate answers and returns a probability for each, whereas a generative model creates free‑form text. Decision models are useful when you need a confidence‑scored answer rather than open‑ended generation.

Can I run d3-ultra on my laptop?

Running the full 27 B version requires a high‑memory GPU. However, the vLLM team may later release quantised or distilled variants that can operate on consumer‑grade hardware.

Is the model compatible with existing Graf by workflows?

Yes. By adding the model to Graf by’s Chat or Graph apps, you can feed multimodal inputs and retrieve probability scores, fitting seamlessly into existing creative and development pipelines.

---

Source: MarkTechPost — vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model