What it is
d3-turbo is the upgraded 27B multimodal decision model of Decision 3.0, from the vLLM Semantic Router project. It was published on Hugging Face on 19 October 2026 under Apache 2.0, nine days after d3, as a separate set of weights fine-tuned from Qwen3.8-27B (26.09 billion parameters including a 0.46B vision encoder). It now serves the free API at decision.vllm-sr.ai, where the name vllm-sr/d3 is an alias for it.
Like d3 it answers choice, yes/no and score questions with a probability for every option, without generating text. It adds Box answers (where an object is in an image or at a moment of a video, with a confidence), whole-video tracking, images placed inside questions and options, and a long-context mode of 131,072 tokens per question, against 32,768 by default.
Its main architectural change is reading the input once per request and answering every question from that single read. Stock vLLM support and quantised builds were still pending when we checked; it runs through Transformers or the Semantic Router's own Engine mode.
Our review
The headline improvement is architectural, not a leaderboard jump. On the community Jev Decision Index, updated on 22 October 2026, d3-turbo v1.1 sits 10th of 229 entries with a balanced skill of 67.31, a hair above d3 at 67.15. Its public-suite score is clearer, 71.62 against d3's 65.45 after the index's own rescore, and the card's estimates of 70.2 on text and 72.9 on vision are the team's, not official board results.
What changes the economics is reading once. According to the card, on one AMD Instinct MI325X, 64 questions about one photo take 867 ms on d3-turbo against 8.26 s on the 9B d3-flash, and 255 questions take 2.12 s against 32.94 s. With a single question the order flips: d3-flash answers in 154 ms, d3-turbo in 302 ms. A moderation pipeline asking a dozen policy questions of every upload is where it pays; a router asking one question per prompt is better served by a smaller sibling.
Box and tracking answers move it beyond classification. The team reports 0.90 accuracy at an overlap threshold of 0.5 on its own Box set and a presence F1 of 0.97 on its tracking clips. Every Box question returns exactly one box even when the object is absent, so the card advises pairing it with a yes/no question, a detail that will catch out anyone wiring it straight into an alerting rule.
Serving is less mature than d3's. The pull request adding d3-turbo's attention pattern to vLLM (#62584) was still open on our checked date, there are no FP8 or 4-bit builds yet (the project tracks an NVFP4 package as an open issue), and the documented Engine path targets AMD ROCm from the router's dev channel. At 26.09 billion BF16 parameters the weights are about 52 GB, which rules out a single 24 GB card until quantised builds land.
Where it earns its place
A trust-and-safety lead at a marketplace checking 50,000 listing photos a day against 20 policy rules
Run every rule against every photo without paying for 20 model calls
Send each photo with all 20 rules as yes/no questions in one request; the card's figures put 16 questions at 426 ms on one MI325X, so 50,000 listings is about six GPU-hours a day. Probabilities per rule let reviewers set separate thresholds.
A computer-vision engineer at a retail-analytics firm
Find whether, and where, a product appears in store camera clips
Ask a yes/no presence question and a track question for the product over each clip at one or two updates a second. The free API accepts one video of up to 60 seconds; production volume means self-hosting the 52 GB of BF16 weights.
An engineering manager already running d3 behind the free API
Understand why answers changed without a code change
The free API's vllm-sr/d3 name now points to d3-turbo, so answers can differ from those recorded before d3-turbo's release on 19 October. Pin a self-hosted checkpoint if reproducibility matters, and do not send sensitive data to the free API, which logs requests to improve the models.
Before you commit
- Setup
- Install Transformers 5.17.0 and load with AutoModel and trust_remote_code, or run vllm-sr serve vllm-sr/d3-turbo --engine --platform rocm from the router's dev channel. The card reports about 2.5 minutes for the first Engine start and 35 seconds after.
- Integrations
- The System One request format shared with Jev, at /v1/systemone on the free API and the Engine; a Transformers system_one() method; a Hugging Face Space. Stock vLLM support is an open pull request.
- 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 API states that requests and responses may be logged to improve the models.
- Cost at scale
- No hosted price. BF16 weights are about 52 GB, so one 80 GB-class GPU per replica; the card publishes no memory peak, though the same-sized d3 peaked at 62 to 65 GiB. Quantised builds are not yet available.
- Limits
- 32,768 tokens per question by default, 131,072 with long_context; over-long questions return max_length_exceeded rather than truncating. The free API caps requests at 64 questions, 16 images of 16 MiB, one 60-second video and a 35-second timeout.
What's good
- 10th of 229 on the community Decision Index (67.31)
- One read answers up to 255 questions: 2.12 s for 255 on one MI325X
- Box and video-tracking answers with confidences
- Apache 2.0 with a free public API to try it
What's not
- Slower than d3-flash for single-question requests
- Not yet supported in stock vLLM; no quantised builds
- Free API logs requests and has no SLA
- Most figures are the team's own, on AMD hardware
Best for
Teams with GPUs who ask many yes/no, choice or location questions of the same text, image or video and want open weights.
Alternatives
d3-turbo (vLLM Semantic Router): common questions
- Is d3-turbo (vLLM Semantic Router) worth it?
- d3-turbo earns its place when one input faces many questions, and costs more than d3-flash when it faces only one. We rate it 4.1 out of 5.
- How much does d3-turbo (vLLM Semantic Router) cost?
- Free weights under Apache 2.0. No paid hosted API; a free public API with logging and no SLA.
- What are the downsides of d3-turbo (vLLM Semantic Router)?
- Slower than d3-flash for single-question requests; not yet supported in stock vLLM; no quantised builds; free API logs requests and has no SLA; most figures are the team's own, on AMD hardware.
- Who is d3-turbo (vLLM Semantic Router) best for?
- Teams with GPUs who ask many yes/no, choice or location questions of the same text, image or video and want open weights.
- How much does d3-turbo (vLLM Semantic Router) cost for a team or in production?
- No hosted price. BF16 weights are about 52 GB, so one 80 GB-class GPU per replica; the card publishes no memory peak, though the same-sized d3 peaked at 62 to 65 GiB. Quantised builds are not yet available.
- Is d3-turbo (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 API states that requests and responses may be logged to improve the models.
- What are the best d3-turbo (vLLM Semantic Router) alternatives?
- The closest alternatives we have reviewed are d3 (vLLM Semantic Router), Perplexity Decider (pplx-decider), Jev, Clef and GLiNER2.5-Decide.
Answers checked Oct 23, 2026.
Make d3-turbo (vLLM Semantic Router)? Show this rating on your site
