What it is
d3 is the 27B multimodal decision model of Decision 3.0, from the vLLM Semantic Router project. It was published on Hugging Face on 10 October 2026 under Apache 2.0, fine-tuned from Qwen3.8-27B (26.09 billion parameters including a 0.46B vision encoder). Nine days later the team released d3-turbo, an upgraded 27B model that now serves the free demo at decision.vllm-sr.ai; d3 remains available as weights.
It takes text or JSON plus several images or videos per request and answers choice, yes/no and score questions with a probability for every option, without generating text. Images are read at up to 1.6 megapixels, videos at 2 frames per second, up to 32 frames.
vLLM's main branch has loaded d3 as-is since 14 October 2026 and answers on a /v1/systemone endpoint, the same request shape as Jev. Official FP8, NVFP4 and int4 (AWQ and GPTQ) builds are published; the 4-bit ones cut the 52.2 GB of weights to about 16 to 17 GB.
Our review
d3 is a solid open decision model that has already been overtaken. On the community Jev Decision Index 0.4.1, updated on 22 October 2026, it sits 11th of 229 entries at 67.15, inside a tie band and just behind its sibling d3-turbo at 67.31. Its card, written against the official 0.3.2 board, claimed first place on text with 66.3; the index and the project have both moved on. Jev 1.14 scores 60.24 on the same community board.
Serving is the reason to look at it. vLLM loads the repository with no conversion and exposes the Jev-style /v1/systemone endpoint, though the support landed on main on 14 October and, in our reading of the repository, is not in the v0.32.0 release, so a nightly build is needed. The card reports a 54.4 ms median for one text question, 290 ms with an image and 500 ms with a 10-second video on one AMD Instinct MI325X, and 27 requests a second with p99 under one second through the Semantic Router. Those are the project's own figures on AMD hardware; the CUDA path had not been run on an NVIDIA GPU when the card was written.
Video input is still rare among decision models. The card reports 77.4 on the Perception Test validation set, an internal evaluation, and the community vision board places d3 ninth of 55 at 71.26. d3-turbo goes further, adding bounding boxes, video tracking and a 131,072-token long-context mode, and it reads the input once however many questions are asked.
Memory is the operational cost. A BF16 replica peaked at 62 to 65 GiB on the MI325X, and vLLM must be started with --max-model-len 32768 or it profiles a 262,144-token pass that does not fit on an 80 GB GPU. The int4 builds bring weights to about 16 GB and keep roughly 97% of d3's answers, within reach of a 24 GB card, but the NVFP4 build needs a Blackwell GPU and the AWQ build runs 1.29 times slower than BF16. The same recipe now applies to d3-turbo, so the main reason to pin d3 is a pipeline already validated against this exact checkpoint.
Where it earns its place
A platform engineer running vLLM on 80 GB GPUs
Add routing and policy checks in front of an LLM gateway without a new runtime
Serve d3 or its FP8 build with --max-model-len 32768 and send each request to /v1/systemone with choice and yes/no questions. The card reports 54 ms for a one-question text decision and 27 requests a second per MI325X; the cost is the GPU, about 62 to 65 GiB per BF16 replica.
A trust-and-safety lead at a marketplace screening 50,000 listing photos a day
Flag listings whose photos contradict the description
Send listing text and photos in one request and ask yes/no and score questions; every answer carries a probability, so review thresholds can be tuned. At 290 ms per image request on one GPU, 50,000 requests is about four GPU-hours a day before batching.
An engineering manager weighing self-hosting against a hosted decision API
Decide whether running an open model is worth it
d3 has no paid endpoint, so the cost is GPUs plus on-call time, and the free demo now runs d3-turbo with no SLA. Teams without GPU operations will spend less staff time on a hosted option such as pplx-decider or Jev.
Before you commit
- Setup
- Install the pinned Transformers 5.17.0 and load with AutoModel and trust_remote_code, or install a vLLM nightly and run vllm serve vllm-sr/d3 --max-model-len 32768. First start with empty caches took 215 s in vLLM (139 s with the length cap) against 40 s for AutoModel.
- Integrations
- vLLM /v1/systemone (nightly), vLLM Semantic Router Engine mode (dev channel), a Transformers pipeline and a bundled d3_server.py. The request format matches Jev's.
- Data and privacy
- Apache 2.0, inherited from Qwen3.8-27B; commercial use allowed with no revenue cap. Self-hosted, inputs stay on your hardware; the free demo's data terms are not published.
- Cost at scale
- No hosted price. BF16 weights are 52.2 GB and a replica peaked at 62 to 65 GiB on one MI325X. The NVFP4 build needs about 18 GB free on a Blackwell GPU; the AWQ int4 build about 18 GB on any GPU, at 1.29 times the latency.
- Limits
- A 32-frame video is about 6K tokens. A top-level instructions field returns 400. Engine mode admits 32 requests in flight per worker and returns 503 for the rest, and eight-question image requests missed a one-second p99 at every load tested.
What's good
- 11th of 229 on the community Decision Index 0.4.1 (67.15)
- Loads in stock vLLM with a Jev-compatible endpoint
- Reads several images and videos per request
- Official FP8, NVFP4 and int4 builds from about 16 GB
What's not
- Superseded by d3-turbo nine days after release
- vLLM support only in nightly builds so far
- Speed figures are the project's own, on AMD MI325X
- Smaller siblings score far lower: d3-flash 54.27, d3-mini 49.97
Best for
Teams already running vLLM who want an open decision model for text, image or video checks on their own GPUs and have validated against this checkpoint.
Alternatives
d3 (vLLM Semantic Router): common questions
- Is d3 (vLLM Semantic Router) worth it?
- Strong, open and easy to serve on vLLM, but nine days after launch its own d3-turbo made it the version to migrate from. We rate it 4.0 out of 5.
- How much does d3 (vLLM Semantic Router) cost?
- Free weights under Apache 2.0. No paid hosted API; the free public demo now serves d3-turbo.
- What are the downsides of d3 (vLLM Semantic Router)?
- Superseded by d3-turbo nine days after release; vLLM support only in nightly builds so far; speed figures are the project's own, on AMD MI325X; smaller siblings score far lower: d3-flash 54.27, d3-mini 49.97.
- Who is d3 (vLLM Semantic Router) best for?
- Teams already running vLLM who want an open decision model for text, image or video checks on their own GPUs and have validated against this checkpoint.
- How much does d3 (vLLM Semantic Router) cost for a team or in production?
- No hosted price. BF16 weights are 52.2 GB and a replica peaked at 62 to 65 GiB on one MI325X. The NVFP4 build needs about 18 GB free on a Blackwell GPU; the AWQ int4 build about 18 GB on any GPU, at 1.29 times the latency.
- Is d3 (vLLM Semantic Router) safe for company data?
- Apache 2.0, inherited from Qwen3.8-27B; commercial use allowed with no revenue cap. Self-hosted, inputs stay on your hardware; the free demo's data terms are not published.
- What are the best d3 (vLLM Semantic Router) alternatives?
- The closest alternatives we have reviewed are Perplexity Decider (pplx-decider), Jev, Clef, GLiNER2.5-Decide and Laya.
Answers checked Oct 22, 2026.
Make d3 (vLLM Semantic Router)? Show this rating on your site
