vLLM Semantic Router's d3 Models Answer Routing Questions About Text, Images and Video in One Call

The team behind vLLM Semantic Router, the request-routing layer in the vLLM project, has released Decision 3.0, a family of open decision models that read text, JSON, images and now video, and answer structured questions with probabilities rather than prose. Six checkpoints are on Hugging Face under the vllm-sr organization, all Apache-2.0: d3-edge, d3-lite, d3-nano, d3-mini, d3-flash and the flagship d3. They were uploaded on October 10, 2026.

A model that never writes a sentence

A decision model is built for the moments in an AI system where nobody needs a paragraph, only a verdict: which team should take this ticket, does this document contain a signature, how urgent is the request. The usual workaround is to prompt a general chat model and parse its reply, which is slow and brittle. The d3 cards describe a different contract.

A caller passes a “state,” which can be plain text or JSON, plus optional images and videos, and a set of named questions. Each question is one of three types: a choice among labeled options, a yes-or-no, or a score on an ordered scale. The model answers every question in a single call and returns a probability for each possible answer. Each question gets its own forward pass over the input, and every question sees all attached media.

That output shape is what makes these models useful in agent pipelines and gateways. A probability can be thresholded, logged and compared over time; a generated sentence has to be interpreted first. The project calls this the “system one” pattern, and the Python entry point is literally model.system_one().

Six sizes on two Qwen bases

The lineup spans a wide range. d3-edge has 0.59 billion parameters, d3-lite 0.85 billion, d3-nano 2.21 billion, d3-mini 4.54 billion and d3-flash 8.39 billion. The flagship d3 has 26.09 billion, including a 0.46-billion-parameter vision encoder. The smaller models are fine-tunes of Qwen3.5 checkpoints, such as Qwen3.5-2B for d3-nano and Qwen3.5-9B for d3-flash, while d3 is built on Qwen3.8-27B.

Every size accepts several PNG, JPEG or WebP images, each read at up to 1.6 megapixels. Video is the headline addition: MP4, WebM, MOV and MKV inputs are sampled at two frames per second, up to 32 frames per video at no more than 0.2 megapixels each. On the Perception Test multiple-choice video benchmark, the team reports 77.4 for d3 and 73.3 for d3-flash, both labeled internal evaluation.

Loading requires transformers 5.17.0 and trust_remote_code=True; an optional flash-linear-attention package speeds up the linear-attention layers.

How the scores stack up

The team measures against the Jev Decision Index, a public board for decision models. Using the official 0.3 evaluation kit on the released weights, the cards report public-suite scores of 65.45 for d3, 57.79 for d3-flash, 54.90 for d3-mini, 42.22 for d3-nano, 36.93 for d3-lite and 28.82 for d3-edge, with all 140,178 public requests answered.

Each card also states the gain over the matching Decision 2.0 model on the same suite: +8.5 for the 27B tier, +11.0 for 9B, +16.4 for 0.8B and +12.1 for 0.6B. On the combined 0.3.1 index table in the d3 card, the flagship scores 64.7, ahead of Perplexity Decider v1.1 at 62.8 and Jev at 60.1. Torchcast Decision 27B comes closest on the public suite, at 65.1 to d3’s 65.5, but trails on the new-domain tasks, 50.8 to 56.6. On the vision board, d3 scores 71.6 to Perplexity Decider’s 70.6.

The card is explicit that d3 figures come from the team’s internal evaluation, while competitor numbers are taken from live board data dated October 10 for text and October 9 for vision.

The public vision results are uneven in useful ways. d3 is near the ceiling on document tasks, with 97.7 on key information extraction and 96.7 on InfographicVQA, but scores 43.9 on Hateful Memes moderation and 33.0 on R-Bench-M. Teams planning to use it as a content-safety gate on images should test that use case directly.

Speed on a single GPU

All models were trained on AMD Instinct MI325X GPUs, and latency is reported on one MI325X handling one request at a time. Median text latency ranges from 6.7 milliseconds for d3-edge to 55 milliseconds for d3. A request with an image takes a median of 273 milliseconds on d3, and a request with a 10-second video 553 milliseconds. d3-flash comes in at 19.1 milliseconds for text.

Those figures put the middle of the range within reach of inline use: a router or guardrail that answers several typed questions about every request before it reaches a larger model, without generating a single token of text.