vLLM Semantic Router has just dropped d3-ultra, a 27B open-weight multimodal decision model built to skip text generation entirely. Instead, it focuses on fast, probabilistic decisions. The model takes in text, images, or video in one pass and spits out calibrated probabilities along with spatial coordinates. Licensed under Apache-2.0, it gives infrastructure teams a quicker, more reliable option than twisting generative LLMs into routers.
Quick Take
- The paradigm shift: d3-ultra does not generate text; it is a pure “decision model” that ingests multimodal data and returns exact probabilities, bounding boxes, or video tracking coordinates.
- Enterprise-grade scale: Featuring 26.09B parameters and a 0.46B vision encoder, it handles a massive 131,072-token context window per question.
- The hardware flex: Benchmarked and served on AMD Instinct MI325X GPUs, signaling a maturing open-source AI stack that is actively optimizing for non-Nvidia silicon.
- Why it matters: It solves the “LLM as a router” bottleneck, allowing developers to classify data, moderate content, and execute visual grounding in milliseconds rather than waiting for token-by-token generation.
What happened
The vLLM Semantic Router organization has officially launched d3-ultra, the flagship 27B parameter model in its new Decision 3.0 family. Available on Hugging Face under an Apache-2.0 license, d3-ultra is engineered specifically for routing, classification, visual grounding, and video analysis. Rather than operating as a chatbot, it uses the “System One” API to answer typed questions—multiple-choice, yes/no, numeric scoring, and spatial localization—based on a single read of the input data.
The model architecture pairs a massive language backbone with a 0.46B parameter vision encoder, allowing it to process complex multimodal workflows natively. For spatial and temporal tasks, it can return precise box coordinates for image localization and track outputs to follow targets through video clips.
Performance data highlights its efficiency at scale. Running on AMD Instinct MI325X hardware, d3-ultra can process a single photo question in 304 milliseconds, with subsequent questions taking just 7.3 milliseconds each. The model has already claimed the #1 spot on both the text and vision boards of the Jev Decision Index.
Why it matters now
The current generative AI pipeline is highly inefficient at structural routing. Developers routinely force heavy generative LLMs to act as decision engines—prompting them to output strict JSON or classify documents—which wastes compute on token generation and risks hallucination. d3-ultra introduces a purpose-built infrastructure layer for the AI ecosystem. By returning calibrated per-answer probabilities in one pass, it allows applications to make deterministic routing decisions at a fraction of the latency.
This release also highlights a critical evolution in AI hardware dynamics. The vLLM project’s explicit collaboration with AMD to train, serve, and benchmark the Decision 3.0 family on MI325X GPUs is a strategic signal. It proves that the open-source community is building high-throughput inference recipes optimized for AMD, providing enterprise data centers with viable silicon alternatives to Nvidia’s dominant H100s.
For developers, the integration of native visual grounding and video tracking inside a single API call drastically simplifies multimodal application architecture. Instead of stringing together separate computer vision models, text classifiers, and orchestration scripts, engineering teams can use one unified model to triage complex, mixed-media inputs at the edge of their infrastructure.
Who is most affected
| Stakeholder | Impact | Why |
|---|---|---|
| ML Platform Engineers | Cheaper, faster routing pipelines | Replaces bulky generative LLMs with one-pass probabilistic decision models, saving compute and latency. |
| Multimodal App Developers | Simplified visual and video logic | Native support for bounding boxes and video target tracking eliminates the need for fragmented, multi-model architectures. |
| AMD GPU Ecosystem | High-profile validation | Showcases MI325X throughput on modern 27B models, boosting confidence in AMD-based AI inference infrastructure. |
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This is an independent i10x analysis built for ML engineers, AI infrastructure operators, and technical founders. It synthesizes official vLLM project announcements, API documentation, GitHub runtime records, and independent benchmark reporting to contextualize the shift from generative models to decision models.
- vllm-sr/d3-ultra · Hugging Face
- vLLM on X: Great work from AMD and the vLLM Semantic Router team on d3-ultra
- vLLM on X: D3 ranked #1 on both Jev Decision Index 0.3.2 boards
- API Docs · Decision 3.0
- vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model
- semantic-router/src/model-runtime/docs/records/d3-ultra.md
i10x Perspective
The release of d3-ultra signals the “unbundling” of the monolithic LLM. As enterprise AI applications mature, the industry is realizing that using a 100B+ generative model to route a customer service ticket or spot an object in a video is an architectural overkill. The future of AI infrastructure relies on a dual-brain system: hyper-fast, probabilistic “decision models” acting as the brain stem to route and classify, while generative models act as the frontal lobe for complex synthesis. Furthermore, vLLM’s aggressive optimization for AMD Instinct hardware proves that the software layer is actively commoditizing the silicon beneath it, slowly loosening the grip of single-vendor hardware lock-in.
