vLLM Semantic Router releases d3, an open decision model that also reads video
vLLM Semantic Router has released Decision 3.0 under Apache 2.0, a family of decision models that answer questions with options and probabilities. The largest, d3, also reads images and videos, and five smaller versions go down to 0.59 billion parameters.
vLLM Semantic Router, the routing project of the open-source AI server vLLM, released the Decision 3.0 family of decision models on Hugging Face on October 10, 2026. Decision models answer questions about a given input by picking options and return a probability for every option. For example, they can tell in one call which team a customer message should go to, whether a receipt is attached and how urgent the request is. The family has six models from 0.59 to 26 billion parameters, all licensed under Apache 2.0, so they can run on your own servers, including for commercial use.
It reads images and videos too
The largest model, d3, is built on Qwen3.8-27B and carries a 0.46-billion-parameter vision encoder. According to the model card, it can read several images (PNG, JPEG, WebP) and videos (MP4, WebM, MOV, MKV) in one request, sampling video at 2 frames per second and up to 32 frames. There are three question types: picking one of several options, answering yes or no, and rating on a scale. All questions about the same input are answered in one call.
The team says d3 ranks first with 66.3 points on the official text board of the Jev Decision Index 0.3.2, which compares decision models, and is 10.4 points ahead of the previous generation's Decision 2.0 27B model. Speed was measured on a single AMD Instinct MI325X: median latency is 54 milliseconds for a text request, 290 milliseconds with an image and 500 milliseconds with a 10-second video.
Smaller sizes and how to run them
The smaller models are based on Qwen3.5: d3-flash has 8.4, d3-mini 4.5, d3-nano 2.2, d3-lite 0.85 and d3-edge 0.59 billion parameters. The models run directly with Hugging Face Transformers. vLLM Semantic Router also serves Decision 3.0 on your own machine through the System One API in Engine mode, which is currently in the project's development channel. GGUF versions of the previous Decision 2.0 models were released on October 8.
The rankings and speed figures were shared by the team, and the model card gives no separate information on supported languages. Before using it on your own text, images or videos, test it with your own samples. For a fuller picture, follow evaluations from independent researchers as well.
Sources: Hugging Face: d3 model card, GitHub: vLLM Semantic Router, Hugging Face: d3-mini model card.
